UV topcoat performance detection and analysis method and device based on image data processing

Abstract

The embodiment of the present invention discloses a UV topcoat performance detection and analysis method and device based on image data processing, which belongs to the field of image analysis technology. The method comprehensively covers the topcoat status information at different stages by acquiring multiple first UV topcoat performance detection images within the upstream and downstream set detection cycles. Based on the specific cycle and image description channel mining performance characterization vector set, it can accurately quantify hardness, wear resistance, glossiness, adhesion and other aspects of performance. In-depth analysis with a smaller set performance analysis task cycle can capture the details of performance changes. The target performance evaluation decision algorithm debugged by a large number of training images is used to make performance evaluation decisions. This multi-dimensional and high-precision detection method helps to timely discover potential defects of UV topcoat, ensure the quality of UV topcoat, improve the overall quality and reliability of products using UV topcoat, reduce product problems caused by poor topcoat performance, and reduce quality control costs in the production process.

Description

Technical Field

[0001] The embodiments of the present invention belong to the field of image analysis technology, and specifically relate to a UV topcoat performance detection and analysis method and device based on image data processing. Background Art

[0002] In many industries, UV topcoats are widely used on the surfaces of various products to provide aesthetics, protection and other functions. However, there are many problems with traditional UV topcoat performance testing methods. The detection of UV topcoat performance is often lacking in comprehensiveness, and it is not possible to effectively monitor it at different stages, which makes it difficult to obtain complete information on topcoat performance changes. When analyzing topcoat performance, there is a lack of effective means to accurately quantify different properties (such as hardness, wear resistance, etc.), making it difficult to gain an in-depth understanding of the details of performance changes. In addition, the detection accuracy is insufficient, and there is a lack of effective evaluation algorithms debugged based on a large amount of data to accurately determine whether the topcoat has defects. These problems may lead to unstable product surface quality, affect the overall product quality, increase the defective rate, and thus increase production costs. Summary of the invention

[0003] The embodiments of the present invention provide a method and device for detecting and analyzing UV topcoat performance based on image data processing, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] The embodiment of the present invention provides a UV topcoat performance detection and analysis method based on image data processing, which is applied to a UV topcoat performance detection and analysis device. The method comprises: obtaining a plurality of first UV topcoat performance detection images within a detection cycle set upstream and downstream of a UV topcoat performance detection image to be analyzed; mining performance characterization vectors of the plurality of first UV topcoat performance detection images based on the set detection cycle, the set performance analysis task cycle and the set image description channel to obtain a set of performance detection characterization vectors to be analyzed for the UV topcoat performance detection images to be analyzed; the set performance analysis task cycle is less than the set detection cycle; the performance detection image to be analyzed is The test characterization vector set includes a hardness performance test characterization vector set, a wear resistance test characterization vector set, a glossiness test characterization vector set and an adhesion test characterization vector set; a performance evaluation decision is made on the performance test characterization vector set to be analyzed through a target performance evaluation decision algorithm to obtain a performance evaluation decision label of whether the UV topcoat performance test image to be analyzed has defects; the target performance evaluation decision algorithm is obtained by debugging an initial performance evaluation decision algorithm based on a priori defect annotations of a number of second UV topcoat performance test images within a test cycle set according to the upstream and downstream of a number of topcoat performance test training images and whether each topcoat performance test training image has defects.

[0005] An embodiment of the present invention provides a UV topcoat performance detection and analysis device, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above-mentioned method.

[0006] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] The embodiment of the present invention comprehensively covers the status information of the topcoat at different stages by acquiring multiple first UV topcoat performance detection images within the upstream and downstream set detection cycles. By mining the performance characterization vector set based on a specific cycle and image description channel, various performance aspects such as hardness, wear resistance, glossiness and adhesion can be accurately quantified. In-depth analysis with a smaller set performance analysis task cycle can capture the details of performance changes. The target performance evaluation decision algorithm debugged by a large number of training images is used to make performance evaluation decisions with high accuracy. This multi-dimensional, high-precision detection method helps to timely discover potential defects of UV topcoat, ensure the quality of UV topcoat, improve the overall quality and reliability of products using UV topcoat, reduce product problems caused by poor topcoat performance, and reduce quality control costs in the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flow chart of a UV topcoat performance detection and analysis method based on image data processing provided by an embodiment of the present invention;

[0009] Figure 2 A schematic structural diagram of a UV topcoat performance detection and analysis device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0010] Figure 1 A UV topcoat performance detection and analysis method based on image data processing is shown, which is applied to a UV topcoat performance detection and analysis device. The method includes the technical solutions recorded in S110-S130.

[0011] S110: Acquire a plurality of first UV topcoat performance detection images within a set detection period upstream and downstream of the UV topcoat performance detection image to be analyzed.

[0012] S120: Based on the set detection cycle, the set performance analysis task cycle and the set image description channel, performance characterization vector mining is performed on the plurality of first UV topcoat performance detection images to obtain a set of performance detection characterization vectors to be analyzed of the UV topcoat performance detection images to be analyzed.

[0013] Among them, the set performance analysis task cycle is smaller than the set detection cycle; the performance detection characterization vector set to be analyzed includes a hardness performance detection characterization vector set, a wear resistance detection characterization vector set, a glossiness detection characterization vector set and an adhesion detection characterization vector set.

[0014] S130: Performing a performance evaluation decision on the performance detection characterization vector set to be analyzed by using a target performance evaluation decision algorithm to obtain a performance evaluation decision label of whether the UV topcoat performance detection image to be analyzed has defects.

[0015] Among them, the target performance evaluation decision algorithm is obtained by debugging the initial performance evaluation decision algorithm based on a priori defect annotations of a number of second UV topcoat performance detection images within a detection cycle set upstream and downstream of a number of topcoat performance detection training images and whether each topcoat performance detection training image has defects.

[0016] In S110, a specific image acquisition device is required to collect the performance test image of the UV topcoat. A high-resolution industrial camera can be used, such as an industrial camera with a resolution of 2048*1536 pixels, which can clearly capture the micro and macro features of the UV topcoat surface. At the same time, in order to ensure the image quality, the lighting system uses a uniform diffuse reflection light source, and the light intensity is set to 500-1000 lux to avoid the influence of shadows and highlight reflections on the image.

[0017] In addition, setting the detection cycle is a key parameter that determines the time range for obtaining the first UV topcoat performance detection image. The determination of this cycle needs to consider many factors, such as the curing speed of the UV topcoat, the expected time scale of performance changes, and the rhythm of the production process. For example, if the curing process of the UV topcoat is basically completed within 1-2 hours, and some performance changes caused by interaction with the environment may occur within the subsequent 12 hours, then the detection cycle can be set to 12-24 hours. This can fully capture the changes in the topcoat performance during the initial curing and subsequent short periods of time.

[0018] The upstream and downstream setting detection cycle in the embodiment of the present invention means obtaining a series of images in chronological order within a specific time period (for example, it can be understood as the preceding time period / post-sequential time period of the setting detection cycle). For example, the upstream detection cycle can be from the start of UV topcoat coating to a certain intermediate time point, and the downstream detection cycle is from the intermediate time point to a later time point. For example, the coating starts at time t=0, the detection cycle is set to 24 hours, and the intermediate time point is t=12 hours, then the upstream detection cycle is 0-12 hours, and the downstream detection cycle is 12-24 hours.

[0019] Next, within this set detection cycle, images are acquired at certain time intervals. For example, if an image is acquired once an hour, 24 first UV topcoat performance detection images can be acquired within a set detection cycle of 24 hours. These images will serve as the basic data for subsequent performance analysis.

[0020] In S120, the key technical point is the mining of the performance characterization vector of the first UV topcoat performance test image and the determination of the performance test characterization vector set to be analyzed.

[0021] 1. Relationship between setting performance analysis task cycle and setting detection cycle

[0022] Importance of periodic relationship: Setting the performance analysis task cycle shorter than the set detection cycle is to more carefully analyze the changes in UV topcoat performance in a shorter time scale. For example, if the detection cycle is set to 24 hours, the performance analysis task cycle can be set to 6 hours. This means that within the total detection cycle of 24 hours, the performance characterization vector will be mined with 6 hours as an analysis unit.

[0023] Impact on performance analysis: This cycle setting can capture the performance change characteristics of UV topcoats at different stages. For example, there may be different performance change patterns in the early curing stage (0-6 hours), mid-term stabilization stage (6-12 hours) and late stage of interaction with the environment (12-24 hours) of UV topcoats. These changes can be analyzed more accurately through smaller performance analysis task cycles.

[0024] 2. Concept and function of setting image description channel

[0025] Channel type: Set the image description channel to include color channel, texture channel, shape channel, etc. The color channel can describe the color characteristics of the UV topcoat. For example, in the RGB color space, the distribution and change of the topcoat color can be obtained by analyzing the values ​​of the red, green, and blue channels of each pixel.

[0026] The texture channel is used to characterize the texture characteristics of the paint surface. The gray level co-occurrence matrix (GLCM) algorithm can be used to construct the texture channel. For example, the co-occurrence probability of the pixel gray value in a specific direction (such as horizontal direction, vertical direction) and a specific distance (such as 1 pixel, 2 pixels) is calculated to obtain parameters reflecting the texture roughness, directionality and other characteristics.

[0027] The shape channel focuses on possible shape defects on the paint surface, such as scratches, bumps, etc. The shape channel can be constructed by determining the boundaries of the paint surface shape using edge detection algorithms, such as the Canny edge detection algorithm.

[0028] Advantages of multi-channel comprehensive analysis: By combining multiple image description channels, the performance of UV topcoats can be more fully described. For example, changes in the color channel may be related to the chemical stability of the topcoat, changes in the texture channel may reflect changes in wear resistance and hardness, and changes in the shape channel are directly related to adhesion and surface integrity.

[0029] 3.3 Performance Characterization Vector Mining Process

[0030] Mining of hardness performance detection characterization vector set: Based on image feature extraction, for the mining of hardness performance detection characterization vector set, features related to hardness are extracted from the image. For example, within the set performance analysis task cycle, observe the changes in the microstructure of the topcoat surface at different time points. Use microscopic image analysis technology, such as using an optical microscope to obtain high-magnification images. For example, within each 6-hour task cycle, select 10 fixed microscopic areas for observation; construct characterization vectors, and calculate parameters such as the deformation coefficient and density change of particles in these areas. For example, take the particle deformation coefficient as one vector element and the density change rate as another element to construct a hardness performance detection characterization vector. If within a task cycle, the particle deformation coefficient changes from the initial 0.1 to 0.15, and the density change rate changes from 0.05 to 0.08, then the hardness performance detection characterization vector within this task cycle can be expressed as [0.05, 0.03] (here 0.05 represents the change in the deformation coefficient, and 0.03 represents the change in the density change rate).

[0031] Mining of wear resistance detection characterization vector set: Analysis of wear area characteristics. In the mining of wear resistance detection characterization vector set, the focus is on the characteristics of the wear area on the surface of the topcoat. By analyzing images at different wear stages within a set performance analysis task cycle. For example, an image segmentation algorithm, such as a threshold-based segmentation algorithm, is used to separate the worn area from the unworn area; the vector elements are determined, and the area change rate of the worn area, the average length and width change of the scratches, etc. are calculated as elements of the characterization vector. For example, within a task cycle, the area of ​​the worn area increases from the initial 10 square millimeters to 15 square millimeters, the area change rate is 0.5, the average length of the scratch increases from 5 millimeters to 7 millimeters, and the average width increases from 0.1 millimeters to 0.15 millimeters. Then the wear resistance detection characterization vector can be expressed as [0.5, 2, 0.05] (here 2 represents the change in the scratch length, and 0.05 represents the change in the scratch width).

[0032] Mining of gloss detection characterization vector set: gloss-related image features, mining of gloss detection characterization vector set is mainly based on the brightness and reflection characteristics of the image. By analyzing the characteristics of the highlight area, average brightness, etc. of the image within the task cycle of the set performance analysis. For example, the image histogram analysis method is used to determine the brightness distribution; vector construction, the change in the highlight area ratio, the change in average brightness, etc. are used as vector elements. For example, within a task cycle, the highlight area ratio changes from 0.2 to 0.25, and the average brightness changes from 100 (measured in a certain brightness unit) to 120, then the gloss detection characterization vector can be expressed as [0.05, 20] (here 0.05 represents the change in the highlight area ratio, and 20 represents the change in average brightness).

[0033] Mining of adhesion detection characterization vector set: Peeling and flaking area analysis. For the mining of adhesion detection characterization vector set, the main focus is on analyzing the peeling and flaking of the topcoat surface. Use image segmentation technology to separate the peeling and flaking areas from the normal areas; define vector elements, and calculate the change in the area ratio of the peeling and flaking areas, the change in edge roughness, etc. as vector elements. For example, within a task cycle, the area ratio of the peeling and flaking areas changes from 0.01 to 0.03, and the edge roughness changes from 1.0 (in a certain roughness measurement unit) to 1.2, then the adhesion detection characterization vector can be expressed as [0.02, 0.2] (here 0.02 represents the change in the area ratio of peeling and flaking, and 0.2 represents the change in edge roughness).

[0034] 4. Formation of the performance detection characterization vector set to be analyzed

[0035] Set construction: The hardness performance detection characterization vector set, wear resistance detection characterization vector set, gloss detection characterization vector set and adhesion detection characterization vector set are combined together to form the performance detection characterization vector set to be analyzed for the UV topcoat performance detection image to be analyzed. This set contains various performance-related vector information within multiple task cycles, and comprehensively describes the performance change characteristics of the UV topcoat within the set detection cycle.

[0036] Based on S120, S130 can make a performance evaluation decision on the performance detection characterization vector set to be analyzed through a target performance evaluation decision algorithm, and obtain a performance evaluation decision label of whether the UV topcoat performance detection image to be analyzed has defects.

[0037] 1. Obtaining the target performance evaluation decision algorithm

[0038] Initial performance evaluation decision algorithm and prior defect annotation: The target performance evaluation decision algorithm is debugged based on a number of second UV topcoat performance test images within the upstream and downstream set test cycles of a number of topcoat performance test training images and a priori defect annotations on whether each topcoat performance test training image has defects. The initial performance evaluation decision algorithm can be a classification algorithm based on machine learning, such as a support vector machine (SVM) algorithm. First, a large number of topcoat performance test training images are collected, and for each training image, a number of second UV topcoat performance test images are obtained within the upstream and downstream set test cycles, and each training image is manually annotated to determine whether there are defects (for example, images with scratches, uneven color, peeling, etc. are marked as defective, and images with intact surfaces and stable performance are marked as non-defective).

[0039] Algorithm debugging process: Use these training images with prior defect annotations and the corresponding second UV topcoat performance test images to debug the initial performance evaluation decision algorithm. Take the performance test characterization vector set of the training image to be analyzed as input, and the annotation of whether there are defects as output. By adjusting the parameters of the algorithm (such as kernel function parameters and penalty factors in SVM), the algorithm can accurately classify the training images. For example, using the cross-validation method, divide the training images into multiple subsets, use some of the subsets as training sets in turn, and the remaining subsets as validation sets, and continuously optimize the performance of the algorithm until a satisfactory classification accuracy rate is achieved, such as an accuracy rate of more than 90%.

[0040] 2. Performance evaluation decision-making process

[0041] Evaluation based on target algorithm: Once the target performance evaluation decision algorithm is obtained, the performance detection characterization vector set of the UV topcoat performance detection image to be analyzed can be input into the algorithm. For example, for a UV topcoat performance detection image to be analyzed, its performance detection characterization vector set to be analyzed is {[0.03, 0.02], [0.3, 1, 0.03], [0.04, 15], [0.01, 0.1]} (corresponding to the characterization vectors of hardness, wear resistance, glossiness, and adhesion, respectively).

[0042] Generation of performance evaluation decision labels: The target performance evaluation decision algorithm calculates based on the input vector set to determine whether the image has defects. If the algorithm output is 1, it means there is a defect; if the output is 0, it means there is no defect. This output is the performance evaluation decision label of whether the UV topcoat performance test image to be analyzed has defects.

[0043] Through the above S110-S130, the UV topcoat performance detection and analysis device can be used to comprehensively and accurately detect and evaluate the performance of UV topcoat, providing strong technical support for the quality control and product optimization of UV topcoat.

[0044] It can be seen that the embodiment of the present invention comprehensively covers the state information of the topcoat at different stages by acquiring multiple first UV topcoat performance detection images within the upstream and downstream set detection cycles. Based on the specific cycle and image description channel mining performance characterization vector set, it can accurately quantify various performance aspects such as hardness, wear resistance, glossiness and adhesion. In-depth analysis with a smaller set performance analysis task cycle can capture the details of performance changes. The target performance evaluation decision algorithm debugged by a large number of training images is used to make performance evaluation decisions with high accuracy. This multi-dimensional, high-precision detection method helps to timely discover potential defects of UV topcoat, ensure the quality of UV topcoat, improve the overall quality and reliability of products using UV topcoat, reduce product problems caused by poor topcoat performance, and reduce quality control costs in the production process.

[0045] In actual application, S120 is one of the core invention points. The performance characterization vector mining of the several first UV topcoat performance detection images based on the set detection cycle, the set performance analysis task cycle and the set image description channel is recorded to obtain the performance detection characterization vector set to be analyzed of the UV topcoat performance detection images to be analyzed, which further includes the following steps: performing image block disassembly on the several first UV topcoat performance detection images according to the set performance analysis task cycle within the set detection cycle to obtain the third UV topcoat performance detection image of each set performance analysis task cycle within the set detection cycle; performing performance characterization vector mining on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed of each set performance analysis task cycle; integrating and interacting the performance detection characterization vectors to be analyzed of several set performance analysis task cycles within the set detection cycle to obtain the performance detection characterization vector set to be analyzed.

[0046] It can be understood that the set image description channel includes one or more of a mechanical deformation visual description channel, a reflection characteristic visual description channel and a boundary texture visual description channel; when the set image description channel includes the mechanical deformation visual description channel, the performance detection representation vector to be analyzed includes a mechanical deformation visual description vector; when the set image description channel includes the reflection characteristic visual description channel, the performance detection representation vector to be analyzed includes a reflection characteristic visual description vector; when the set image description channel includes the boundary texture visual description channel, the performance detection representation vector to be analyzed includes a boundary texture visual description vector.

[0047] For S120, when testing and analyzing the performance of UV topcoat, the setting of the test cycle determines a relatively long time range to observe the changes in the performance of the topcoat, and the setting of the performance analysis task cycle is a more detailed analysis unit divided within this long period. The image block disassembly is to subdivide the first UV topcoat performance test image according to the set performance analysis task cycle, so as to more accurately analyze the topcoat performance characteristics of each stage.

[0048] For example, the detection cycle is set to 24 hours, and the performance analysis task cycle is set to 6 hours. This means that within the 24 hours, every 6 hours is used as an independent analysis unit. If 24 first UV topcoat performance detection images are acquired within the 24 hours (for example, 1 image is acquired every hour), then for each 6-hour set performance analysis task cycle, 6 first UV topcoat performance detection images will be included.

[0049] The specific operation of image block disassembly is illustrated with an exemplary numerical example, for example, the size of each first UV topcoat performance detection image is 512*512 pixels. For the image block disassembly of each set performance analysis task cycle, the corresponding first UV topcoat performance detection images can be combined together in chronological order. For example, the image block corresponding to the first 6-hour set performance analysis task cycle is composed of the 1st to 6th first UV topcoat performance detection images, and the image block corresponding to the second 6-hour set performance analysis task cycle is composed of the 7th to 12th first UV topcoat performance detection images, and so on. These combined image blocks become the third UV topcoat performance detection image of each set performance analysis task cycle.

[0050] Next, based on the set image description channel, the performance characterization vector of the third UV topcoat performance detection image of each set performance analysis task cycle is mined to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, wherein the set image description channel includes the following three channels.

[0051] 1) Set up the image description channel - mechanical deformation visual description channel

[0052] Concept of the mechanical deformation visual description channel: The mechanical deformation visual description channel is mainly used to analyze the microscopic and macroscopic deformation of the surface of UV topcoat when it is subjected to external force or internal stress. This deformation may be related to the mechanical properties of the UV topcoat, such as hardness and flexibility.

[0053] Mining of visual description vectors of mechanical deformation: First, for the third UV topcoat performance test image of each set performance analysis task cycle, an image analysis algorithm is used to detect the deformation characteristics of the surface. For example, an algorithm based on the optical flow method can be used to calculate the displacement vector of each pixel in the image to represent the deformation of the surface. For example, within a set performance analysis task cycle, the image size is 512*512 pixels, and for each pixel (x, y), the displacement vector calculated by the optical flow method is (u(x, y), v(x, y)).

[0054] Then, construct the mechanical deformation visual description vector. The average displacement vector of the entire image, the standard deviation of the displacement vector, etc. can be used as vector elements. Assume that the average displacement vector is (overline{u}, overlap{v}), the standard deviation of the displacement vector is sigma_u and sigma_v, then the mechanical deformation visual description vector can be expressed as [overline{u}, overlap{v}, sigma_u, sigma_v]. For example, if within a certain set performance analysis task cycle, it is calculated that overlap{u}=0.1 pixel, overlap{v}=0.05 pixel, sigma_u=0.03 pixel, sigma_v=0.02 pixel, then the mechanical deformation visual description vector is [0.1, 0.05, 0.03, 0.02].

[0055] 2) Set the image description channel - reflection characteristic visual description channel

[0056] Concept of the Reflectance Visual Description Channel: The Reflectance Visual Description Channel focuses on the reflectance characteristics of the UV topcoat surface, which is closely related to the optical properties of the topcoat, such as glossiness. Topcoat surfaces with different glossiness will have different reflection patterns when exposed to light.

[0057] Mining of reflection characteristic visual description vector: For each third UV topcoat performance detection image of the set performance analysis task cycle, calculate its reflection-related features. The brightness distribution of the image can be analyzed based on the grayscale histogram, because brightness has a certain correlation with reflectivity. Assume that the grayscale value range of the image is [0, 255], and calculate the grayscale histogram H (i), where i = 0, 1, cdots, 255.

[0058] When constructing the visual description vector of the reflectance characteristics, some statistical features of the histogram can be used as vector elements. For example, the percentage of pixels in the highlight area (area with grayscale values ​​greater than 200) P_{high}, the average grayscale value overline{G}, etc. Assume P_{high}=0.1, overline{G}=120, then the visual description vector of the reflectance characteristics can be expressed as [0.1, 120].

[0059] 3) Set the image description channel-boundary texture visual description channel

[0060] Concept of boundary texture visual description channel: Boundary texture visual description channel is mainly used to describe the texture characteristics of UV topcoat surface, especially the texture of the boundary part, which is related to the adhesion, wear resistance and other properties of the topcoat. For example, the roughness and texture direction of the surface texture will affect the adhesion between the topcoat and the substrate and the wear resistance during the friction process.

[0061] Mining of boundary texture visual description vectors: For the third UV topcoat performance detection image of each set performance analysis task cycle, the gray level co-occurrence matrix (GLCM) algorithm is used to analyze the texture features. For an image, let its gray level co-occurrence matrix be P (i, j, d, theta), where i, j represent the gray value, d represents the pixel spacing, and theta represents the direction (for example, theta = 0^{circ}, 45^{circ}, 90^{circ}, 135^{circ}).

[0062] When constructing the boundary texture visual description vector, some parameters that can represent the texture features can be selected as vector elements. For example, contrast C = sum_{i, j} (ij) ^ 2 P (i, j, d, theta), correlation R = frac {sum_{i, j} (i-overline{i}) (j-overline{j}) P (i, j, d, theta)} {sigma_isigma_j} (where overline{i}, overline{j} are the means of i and j respectively, and sigma_i, sigma_j are the standard deviations of i and j respectively). For example, in a certain set performance analysis task cycle, C = 10, R = 0.5 is calculated, then the boundary texture visual description vector can be expressed as [10, 0.5].

[0063] Based on the above content, the performance detection characterization vectors to be analyzed of several set performance analysis task cycles within the set detection cycle can be integrated and interacted to obtain the performance detection characterization vector set to be analyzed

[0064] 1) Necessity of integrated interaction: After obtaining the performance detection characterization vectors to be analyzed for each set performance analysis task cycle, these vectors can only reflect part of the performance characteristics within each task cycle. Through integrated interaction, the information of different task cycles can be integrated to form a vector set that comprehensively describes the performance changes of UV topcoat throughout the set detection cycle. This helps to evaluate the performance of UV topcoat more comprehensively and accurately, because the performance of UV topcoat may have different change trends at different stages, and integrated interaction can capture this comprehensive information;

[0065] 2) Specific methods of integrated interaction: For example, there are n set performance analysis task cycles in the set detection cycle, and the performance detection characterization vectors to be analyzed obtained in each task cycle are vec{v}_1, vec{v}_2, cdots, vec{v}_n respectively. For the mechanical deformation visual description vector, for example, the vector in each task cycle is [overline{u}_k, overline{v}_k, sigma_{u, k}, sigma_{v, k}] (k=1, 2, cdots, n), these vectors can be combined in sequence into a new vector set. An exemplary combination method is to arrange the corresponding elements in each task cycle vector in sequence to form a longer vector. For example, the new set of visual description vectors of mechanical deformation can be expressed as [overline{u}_1,overline{v}_1,sigma_{u,1},sigma_{v,1},overline{u}_2,overline{v}_2,sigma_{u,2},sigma_{v,2},cdots,overline{u}_n,overline{v}_n,sigma_{u,n},sigma_{v,n}].

[0066] Similar methods are used to integrate and interact with the reflection characteristic visual description vector and the boundary texture visual description vector. The final performance detection characterization vector set to be analyzed includes the mechanical deformation visual description vector set, the reflection characteristic visual description vector set and the boundary texture visual description vector set, which comprehensively reflects the changes in the hardness, glossiness, adhesion and other aspects of the UV topcoat performance within the set detection cycle.

[0067] The above-mentioned embodiment can analyze the performance changes of UV topcoat at different stages in a more detailed manner by disassembling image blocks according to the set performance analysis task cycle within the set detection cycle. Based on the mining of performance characterization vectors based on a variety of set image description channels, the performance of UV topcoat can be accurately characterized from multiple aspects such as mechanical deformation, reflection characteristics and boundary texture, and quantitative vector elements are obtained through specific algorithms such as optical flow method, grayscale histogram and grayscale co-occurrence matrix, thereby improving the accuracy of performance analysis. The performance detection characterization vectors to be analyzed in different task cycles are integrated and interacted, and the comprehensive information within the entire set detection cycle is integrated to form a comprehensive set of performance detection characterization vectors to be analyzed. This method helps to more accurately evaluate the performance of UV topcoat, discover potential performance problems in a timely manner, ensure the quality of UV topcoat, improve product reliability, reduce the risk of product defects caused by poor topcoat performance, and also provide strong data support for the optimization of UV topcoat production process.

[0068] In an optional example, when the set image description channel includes the mechanical deformation visual description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; performing feature mapping and quantization encoding on the regional grayscale comparison detection data to obtain the first quantization encoding data for each set performance analysis task cycle; performing particle deformation feature mining and quantization encoding on the particle deformation detection data to obtain the second quantization encoding data for each set performance analysis task cycle; performing fusion processing on the first quantization encoding data of each set performance analysis task cycle and the second quantization encoding data of each set performance analysis task cycle to obtain the mechanical deformation visual description vector for each set performance analysis task cycle.

[0069] This example focuses on the introduction of the performance characterization vector mining technology solution when the image description channel is set as the mechanical deformation visual description channel.

[0070] (I) Determine the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle

[0071] 1) Determination of particle deformation detection data

[0072] When analyzing the performance of UV topcoat, particle deformation detection data is an important basis for reflecting the mechanical properties of topcoat. For each third UV topcoat performance detection image of a set performance analysis task cycle, it is first necessary to determine the particle structure in the image. This can be achieved through an image segmentation algorithm, such as a threshold-based segmentation algorithm. For example, the third UV topcoat performance detection image is I (x, y), where (x, y) is the image pixel coordinate, and a suitable grayscale threshold T is set. When I (x, y)>T, the pixel is judged to belong to the particle part. After the particle is determined, the deformation data of the particle is calculated. For example, a template matching algorithm can be used to track the position and shape changes of particles at different times (i.e., images acquired at different times within the set performance analysis task cycle). Assume that the contour coordinates of the particle at the initial moment are {(x_1, y_1), (x_2, y_2), cdots, (x_n, y_n)}. After a period of time, the contour coordinates of the particle become {(x_1', y_1'), (x_2', y_2'), cdots, (x_m', y_m')}. The deformation of the particle can be characterized by calculating the change in the contour coordinates, such as calculating the average coordinate offset overline{Deltax}=frac{1}{n}sum_{i=1}^{n}(x_i'-x_i) and overline{Deltay}=frac{1}{n}sum_{i=1}^{n}(y_i'-y_i), as well as the contour area change rate DeltaS=frac{S'-S}{S}, where S is the initial particle area and S' is the changed particle area. These data constitute the particle deformation detection data;

[0073] 2) Determination of regional grayscale comparison test data

[0074] Regional grayscale comparison test data helps to analyze the mechanical deformation of the topcoat from the perspective of grayscale changes. The third UV topcoat performance test image is divided into several small areas, for example, into MtimesN rectangular areas of equal size, each area is ktimesl pixels in size. For each area R_{ij} (i=1, 2, cdots, M, j=1, 2, cdots, N), calculate its average grayscale value overline{G}_{ij}.

[0075] Then, in the images acquired at different times within the set performance analysis task cycle, the average grayscale value changes of the same area are compared. Assuming the average grayscale value at the initial moment is overline{G}_{ij}^0, and the average grayscale value after the set performance analysis task cycle is overline{G}_{ij}^1, the regional grayscale comparison detection data can be expressed as: {Deltaoverline{G}_{ij}=overline{G}_{ij}^1-overline{G}_{ij}^0}.

[0076] (II) Feature mapping and quantization encoding are performed on the regional grayscale comparison detection data to obtain the first quantization encoding data of each set performance analysis task cycle

[0077] 1) Feature mapping: The purpose of feature mapping the regional grayscale comparison detection data is to convert the information of grayscale value changes into a form that is easier to analyze and process. An exemplary feature mapping method is to use a linear mapping function. Suppose the regional grayscale comparison detection data is Deltaoverline{G}_{ij}, and define the mapping function f(x)=ax+b, where a and b are parameters determined according to actual conditions. For example, when you want to map the grayscale change range from [-50, 50] to [0, 100], you can choose a=1, b=50, then the mapped eigenvalue is f(Deltaoverline{G}_{ij})=Deltaoverline{G}_{ij}+50. This feature mapping can highlight the characteristics of the grayscale comparison detection data and normalize it to a specific range, which is convenient for subsequent quantization encoding;

[0078] 2) Quantization coding: Quantization coding is the process of converting the result after feature mapping into digital coding. The uniform quantization method can be used to divide the range of feature values ​​after mapping into Q quantization intervals. For example, [0, 100] is divided into 10 quantization intervals, and the length of each interval is 10. For each regional grayscale comparison detection data f (Deltaoverline{G}_{ij}) after feature mapping, it is encoded according to the quantization interval in which it is located. For example, f (Deltaoverline{G}_{ij}) = 35, then its code is 3 (because it is in the third quantization interval, counting from 0). In this way, the first quantization coding data of each set performance analysis task cycle is obtained.

[0079] (III) Mining and quantizing the particle deformation characteristics of the particle deformation detection data to obtain the second quantized coded data for each set performance analysis task cycle

[0080] 1) Mining particle deformation features: In the process of mining particle deformation features, in addition to the average coordinate offset and area change rate mentioned above, the deformation direction of the particle can also be calculated. For example, the deformation direction angle is calculated based on the average coordinate offset: theta=arctan(frac{overline{Deltay}}{overline{Deltax}}). At the same time, the deformation concentration of the particle can be calculated, which is expressed by calculating the standard deviation sigma_d of the distance change from each point on the particle contour to the centroid (which can be calculated based on the contour coordinates). These parameters together constitute the particle deformation characteristics. For example, for a particle, it is calculated that overline{Deltax}=2 pixels, overline{Deltay}=3 pixels, DeltaS=0.2, theta=56.31^{circ}, sigma_d=0.5 pixels, these data are the particle deformation feature data;

[0081] 2) Quantization coding: The quantization coding method is also used to process the particle deformation feature data. Different quantization strategies can be used for different feature parameters. For example, for coordinate offset and area change rate, a uniform quantization method similar to the regional grayscale comparison detection data can be used. Suppose the quantization interval of the coordinate offset is [-10, 10], divided into 5 intervals, each interval length is 4; the quantization interval of the area change rate is [-0.5, 0.5], divided into 10 intervals, each interval length is 0.1. For the direction angle theta, since its range is [0, 360^{circ}], it can be divided into 12 quantization intervals, each interval is 30^{circ}. For the standard deviation sigma_d, according to its actual value range, for example [0, 2], it is divided into 4 intervals, each interval is 0.5. Then encode according to the quantization interval where each feature parameter is located, so as to obtain the second quantization encoding data for each set performance analysis task cycle.

[0082] (iv) performing fusion processing on the first quantized coded data of each set performance analysis task cycle and the second quantized coded data of each set performance analysis task cycle to obtain a mechanical deformation visual description vector of each set performance analysis task cycle

[0083] 1) The significance of fusion processing: The purpose of fusing the first quantized coded data and the second quantized coded data is to combine the information obtained from the two aspects of regional grayscale comparison and particle deformation to form a vector that fully describes the visual characteristics of mechanical deformation. Such a vector can more accurately reflect the mechanical deformation of the UV topcoat within the set performance analysis task cycle;

[0084] 2) Fusion processing method: An exemplary fusion method is to arrange the first quantized coded data and the second quantized coded data in a certain order to form a vector. For example, if the first quantized coded data is [q_1, q_2, cdots, q_p] and the second quantized coded data is [r_1, r_2, cdots, r_q], the mechanical deformation visual description vector can be expressed as [q_1, q_2, cdots, q_p, r_1, r_2, cdots, r_q]. The order here can be determined according to actual needs, for example, the codes related to the regional grayscale comparison detection data are arranged first, and then the codes related to the particle deformation detection data are arranged.

[0085] It can be understood that when the image description channel is set as the mechanical deformation visual description channel, the mechanical deformation information of the topcoat is obtained from different angles by determining the particle deformation detection data and the regional grayscale comparison detection data. The data is processed using feature mapping and quantization coding, making the data easier to analyze and consistent. In the particle deformation feature mining, a variety of feature parameters are considered, such as coordinate offset, area change rate, deformation direction and deformation concentration, etc., which enriches the description of particle deformation. The first quantization coding data and the second quantization coding data are integrated to form a mechanical deformation visual description vector, which comprehensively integrates information from different aspects. This method improves the accuracy and comprehensiveness of the mechanical property detection of UV topcoat, and can more carefully capture the mechanical deformation characteristics of the topcoat within the set performance analysis task cycle, providing a reliable basis for accurately evaluating the performance of UV topcoat, and helping to improve the effectiveness of UV topcoat performance detection technology based on image analysis.

[0086] In another optional example, when the set image description channel includes the reflective characteristic visual description channel, and the reflective characteristic visual description channel includes a brightness distribution description channel and a spot state description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the gloss evaluation data and gloss area heat map under the brightness distribution task keyword corresponding to the brightness distribution description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; performing performance characterization vector mining on the gloss evaluation data and gloss area heat map under the brightness distribution task keyword The method comprises the following steps: performing feature mapping and quantization coding to obtain third quantization coding data of each set performance analysis task cycle; determining gloss evaluation data and gloss area heat map under the spot state task keywords corresponding to the spot state description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; performing feature mapping and quantization coding on the gloss evaluation data and gloss area heat map under the spot state task keywords to obtain fourth quantization coding data of each set performance analysis task cycle; performing fusion processing on the third quantization coding data of each set performance analysis task cycle and the fourth quantization coding data of each set performance analysis task cycle to obtain the reflection characteristic visual description vector of each set performance analysis task cycle.

[0087] When the image description channel is set as a reflectance characteristic visual description channel (including a brightness distribution description channel and a light spot state description channel), the technical solution for mining the above-mentioned performance characterization vector is described as follows.

[0088] (I) Determine the gloss evaluation data and gloss area heat map under the brightness distribution task keyword corresponding to the brightness distribution description channel in the third UV topcoat performance detection image of each set performance analysis task cycle

[0089] 1) Determination of gloss evaluation data: When analyzing the reflective characteristics of UV topcoat, for the gloss evaluation data under the brightness distribution description channel, we must first understand the relationship between gloss and brightness distribution. Brightness distribution reflects the reflection and scattering of light on the surface of the UV topcoat, which directly affects the gloss of the topcoat. One method of calculating gloss evaluation data is to use the grayscale weighted average method. Let the third UV topcoat performance detection image be I (x, y), where (x, y) is the image pixel coordinate. Define the weight function w (x, y), for example, the weight can be set according to the distance from the pixel to the center of the image, and the closer the distance to the center, the greater the weight. Then the gloss evaluation data G_{eval} can be calculated by the following formula: [G_{eval}=frac{sum_{x, y}I(x, y)w(x, y)}{sum_{x, y}w(x, y)}]. Here, the numerator represents the total brightness of the image after weighting according to the weight, and the denominator is the sum of the weights. For example, for a 512*512 pixel image, if the weight function is set to a Gaussian function with the center of the image as the center, the weight of the central pixel is 1, and the weight of the edge pixel gradually decreases to 0.1. The calculated gloss evaluation data may be a value between 0-255, and the higher the value, the higher the overall glossiness;

[0090] 2) Determination of gloss area heat map: The gloss area heat map can intuitively show the gloss difference of different areas on the UV topcoat surface. First, the image is divided into several small areas, for example, into MtimesN rectangular areas of equal size, and the size of each area is ktimesl pixels.

[0091] For each region R_{ij} (i=1, 2, cdots, M, j=1, 2, cdots, N), calculate its average gloss. The calculation method is similar to the above gloss evaluation data, but only for the pixels in the region. Let the average gloss of the region be G_{ij}, then a matrix of MtimesN can be constructed to represent the gloss region heat map, where the value of the matrix element (i, j) is G_{ij}. For example, if the image is divided into 16*16 regions, each region is 32*32 pixels, the average gloss of each region is calculated, thereby constructing a gloss region heat map.

[0092] (ii) Feature mapping and quantization encoding of the gloss evaluation data and gloss area heat map under the brightness distribution task keyword to obtain the third quantization encoding data for each set performance analysis task cycle

[0093] 1) Feature mapping: For the feature mapping of gloss evaluation data, for example, the original gloss evaluation data range is [G_{min}, G_{max}], such as [0, 255]. A linear mapping function can be used to map it to a new range, such as [0, 100]. Suppose the mapping function is f(x)=frac{(x-G_{min})}{(G_{max}-G_{min})}times100. For example, if the original gloss evaluation data G_{eval}=128, the mapped eigenvalue is f(128)=frac{(128-0)}{(255-0)}times100=50. For the feature mapping of the gloss area heat map, since the heat map is a matrix, the same linear mapping operation can be performed on each element in the matrix. In this way, the gloss value of each area can be mapped to a new range that is easier to compare and analyze;

[0094] 2) Quantization encoding: When quantizing and encoding the gloss evaluation data, a uniform quantization method is used. The range of the mapped gloss evaluation data, such as [0, 100], is divided into Q quantization intervals, such as 10 intervals, each of which is 10 in length. Encode according to the interval in which the value of the gloss evaluation data is mapped. For example, if the mapped value is 35, it is encoded as 3. For the quantization encoding of the gloss area heat map, since it is a matrix, each element in the matrix can be encoded according to the same quantization interval. For example, for a 16*16 gloss area heat map matrix, after quantizing and encoding each element, a new 16*16 matrix composed of encoded values ​​is obtained, which is the result of the quantization encoding of the gloss area heat map. Combining the encoding of the gloss evaluation data and the quantization encoding results of the gloss area heat map, the third quantization encoding data of each set performance analysis task cycle is obtained.

[0095] (III) Determine the gloss evaluation data and gloss area heat map under the light spot state task keyword corresponding to the light spot state description channel in the third UV topcoat performance detection image of each set performance analysis task cycle

[0096] 1) Gloss evaluation data under the light spot state task keyword: The light spot state is also closely related to the gloss of the UV topcoat. For the gloss evaluation data under the light spot state description channel, it is necessary to pay attention to the intensity and distribution of the light spot on the surface of the topcoat. The light spot center brightness detection algorithm can be used to calculate the gloss evaluation data. First, the light spot area is identified by an image segmentation algorithm. For example, a threshold-based segmentation algorithm is used to determine the light spot area when the pixel brightness is higher than a certain threshold. Suppose the pixels in the light spot area are {(x_1, y_1), (x_2, y_2), cdots, (x_n, y_n)}, then the gloss evaluation data G_{spot} under the light spot state task keyword can be obtained by calculating the average brightness of the light spot area: [G_{spot}=frac{1}{n}sum_{i=1}^{n}I(x_i, y_i)]. For example, for a UV topcoat performance test image containing light spots, after the light spot area is identified, the average brightness of the light spot area is calculated to be 180, which is the gloss evaluation data under the light spot state task keyword;

[0097] 2) Glossy area heat map under the light spot state task keyword: The method of constructing the glossy area heat map under the light spot state task keyword is similar to that under the brightness distribution description channel. Divide the image into several small areas, and then determine whether it contains a light spot and the intensity of the light spot for each area. For each area R_{ij}, if the area contains a light spot, calculate the average brightness ratio P_{ij} of the light spot in the area, that is, the ratio of the total brightness of the light spot pixel to the total pixel brightness of the area. Construct a matrix of MtimesN, where the value of the matrix element (i, j) is P_{ij}. This matrix is ​​the glossy area heat map under the light spot state task keyword.

[0098] (IV) Feature mapping and quantization encoding of the gloss evaluation data and gloss area heat map under the light spot state task keywords are performed to obtain the fourth quantization encoding data of each set performance analysis task cycle

[0099] 1) Feature mapping: For the feature mapping of the gloss evaluation data under the keyword of the light spot state task, a linear mapping function can also be used. For example, its original range is [G_{spot, min}, G_{spot, max}], such as [0, 255], mapped to a new range such as [0, 100], the mapping function is f_{spot}(x)=frac{(x-G_{spot, min})}{(G_{spot, max}-G_{spot, min})}times100. For the feature mapping of the gloss area heat map under the keyword of the light spot state task, the same linear mapping operation is performed on each element in the matrix to map it to a range that is convenient for analysis;

[0100] 2) Quantization coding: For the quantization coding of the gloss evaluation data under the light spot state task keyword, a uniform quantization method is used. The mapped range is divided into several quantization intervals, such as 10 intervals, each of which is 10 in length. Encode according to the interval where the value of the mapped gloss evaluation data is located. For the quantization coding of the gloss area heat map under the light spot state task keyword, each element in the matrix is ​​encoded according to the same quantization interval to obtain a new matrix composed of encoded values. This matrix, together with the encoding of the gloss evaluation data, constitutes the fourth quantization encoding data of each set performance analysis task cycle.

[0101] (V) Fusing the third quantized coded data of each set performance analysis task cycle with the fourth quantized coded data of each set performance analysis task cycle to obtain a reflection characteristic visual description vector of each set performance analysis task cycle

[0102] 1) The significance of fusion processing: The purpose of fusing the third quantized coded data and the fourth quantized coded data is to combine the reflection characteristic information obtained from the brightness distribution and the light spot state to form a vector that fully describes the visual characteristics of the reflection characteristics. This can more accurately reflect the reflection characteristics of the UV topcoat within the set performance analysis task cycle;

[0103] 2) Fusion processing method: An exemplary fusion method is to arrange the third quantized coded data and the fourth quantized coded data in a certain order to form a vector. For example, if the third quantized coded data is [q_1, q_2, cdots, q_p] and the fourth quantized coded data is [r_1, r_2, cdots, r_q], then the reflectance characteristic visual description vector can be expressed as [q_1, q_2, cdots, q_p, r_1, r_2, cdots, r_q]. The order here can be determined according to actual needs, for example, the codes related to brightness distribution are arranged first, and then the codes related to light spot state are arranged.

[0104] It can be seen that when this embodiment sets the image description channel as the reflective characteristic visual description channel, the reflective characteristics of the UV topcoat are comprehensively analyzed from multiple angles by respectively determining the gloss evaluation data and gloss area heat map under the brightness distribution and spot state. In the process of feature mapping and quantization encoding, a reasonable algorithm is used to process the data to make the data standardized and easy to analyze. By fusing the quantized coded data from different aspects to form a reflective characteristic visual description vector, more reflective characteristic information is integrated. This method improves the accuracy and comprehensiveness of the detection of the reflective characteristics of the UV topcoat, and can more carefully capture the changes in the reflective characteristics of the topcoat within the set performance analysis task cycle, providing a reliable basis for accurately evaluating the performance of the UV topcoat, and helping to improve the effectiveness of the entire UV topcoat performance detection technology.

[0105] In another optional example, when the set image description channel includes the boundary texture visual description channel, and the boundary texture visual description channel includes a peeling state visual description channel and a flaking state visual description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the peeling area image block and the peeling area image thermal features corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; The thermal features of the peeling area image are feature mapped and quantized to obtain the fifth quantized coded data of each set performance analysis task cycle; the peeling area image block corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle is determined; the peeling area image block is feature mapped and quantized to obtain the sixth quantized coded data of each set performance analysis task cycle; the fifth quantized coded data of each set performance analysis task cycle and the sixth quantized coded data of each set performance analysis task cycle are fused to obtain the boundary texture visual description vector of each set performance analysis task cycle.

[0106] Furthermore, when the image description channel is set as a boundary texture visual description channel (including a peeling state visual description channel and a flaking state visual description channel), the technical solution for mining the above-mentioned performance characterization vector is introduced as follows.

[0107] (I) Determine the peeling area image block and the peeling area image thermal characteristics corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle

[0108] 1) Determination of the image block of the peeling area: When analyzing the peeling state of the UV topcoat, it is first necessary to accurately identify the peeling area. For the third UV topcoat performance detection image, an image segmentation algorithm based on texture difference can be used to determine the image block of the peeling area. For example, the local texture features of the image, such as the gray level co-occurrence matrix (GLCM) features, are calculated. Let the image be I (x, y), and for each pixel (x, y), its gray level co-occurrence matrix P_d (i, j) in a specific direction (such as horizontal and vertical directions) and a specific distance (such as 1 pixel) is calculated, where i and j represent gray values. Then, the peeling area is identified based on the texture features that the peeling area usually has, such as the specific range of parameters such as contrast and correlation in the gray level co-occurrence matrix. For example, through analysis, it is found that the range of the gray level co-occurrence matrix contrast C of the peeling area is between [10, 20], and the correlation R is between [0.1, 0.3]. When the GLCM features of a certain area in the image meet this range, the area is determined as a peeling area, thereby determining the image block of the peeling area. For example, in a 512*512 pixel image, a 100*100 pixel peeling area image block is identified;

[0109] 2) Determination of thermal characteristics of the peeling area image: The thermal characteristics of the peeling area image can reflect the heat distribution of the peeling area in the image. The "heat" here can be understood as a quantitative representation of the degree of peeling.

[0110] One calculation method is based on the grayscale value change rate of pixels in the peeling area image block. Assume that the peeling area image block is I_{peel}(x, y), where (x, y) is the pixel coordinate in the peeling area image block. Calculate the grayscale value change rate of each pixel DeltaG(x, y)=frac{I_{peel}(x, y)-I_{peel}^0(x, y)}{I_{peel}^0(x, y)}, where I_{peel}^0(x, y) is the grayscale value of the corresponding pixel in the peeling area image block at the initial moment (for example, there is an initial reference image at the beginning of the performance analysis task cycle). Then, the peeling area image block is divided into several small areas, for example, divided into MtimesN small areas of equal size, and the size of each small area is ktimesl pixels. For each small area R_{ij} (i=1, 2, cdots, M, j=1, 2, cdots, N), calculate its average gray value change rate: overline{DeltaG}_{ij}=frac{1}{ktimesl}sum_{(x, y)inR_{ij}}DeltaG(x, y). Use these average gray value change rates to construct a MtimesN matrix, which is the thermal feature matrix of the peeling area image. For example, divide the 100*100 pixel peeling area image block into 10*10 small areas, each of which is 10*10 pixels. The average gray value change rate of each small area is calculated, thereby constructing the thermal feature matrix of the peeling area image.

[0111] (ii) Feature mapping and quantization encoding are performed on the peeling area image blocks and the peeling area image thermal features to obtain the fifth quantization encoding data of each set performance analysis task cycle

[0112] 1) Feature mapping: For the feature mapping of the peeling area image block, the size and shape characteristics of the peeling area image block are first considered. For example, the area A and perimeter P of the peeling area image block can be used as features for mapping. Suppose the original area range is [A_{min}, A_{max}], for example [100, 10000], and it is mapped to the new range [0, 100] using the linear mapping function f_A(A)=frac{(A-A_{min})}{(A_{max}-A_{min})}times100. For the perimeter P, a similar linear mapping function f_P(P) is also used. For the feature mapping of the thermal feature matrix of the peeling area image, each element in the matrix is ​​linearly mapped. For example, the original average gray value change rate range is [DeltaG_{min}, DeltaG_{max}], such as [-0.5, 0.5], mapped to the range of [0, 100], the mapping function is f_{DeltaG}(DeltaG)=frac{(DeltaG-DeltaG_{min})}{(DeltaG_{max}-DeltaG_{min})}times100;

[0113] 2) Quantization coding: When quantizing and coding the features of the peeling area image block, a uniform quantization method is used for the mapped area and perimeter features. For example, the mapped area range [0, 100] is divided into 10 quantization intervals, each interval length is 10; the mapped perimeter range is also divided into 10 quantization intervals. Encoding is performed according to the intervals in which the values ​​of the area and perimeter are mapped. For example, if the mapped value of the area is 35 and the mapped value of the perimeter is 45, the area is coded as 3 and the perimeter is coded as 4. For the quantization coding of the thermal feature matrix of the peeling area image, each element in the matrix is ​​coded according to the same quantization interval (for example, divided into 10 quantization intervals). For example, for a 10*10 thermal feature matrix of the peeling area image, after each element is quantized and coded, a new 10*10 matrix composed of coded values ​​is obtained. The feature coding of the peeling area image block and the coding of the thermal feature matrix of the peeling area image are combined to obtain the fifth quantization coding data of each set performance analysis task cycle.

[0114] (III) Determine the peeling area image block corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle

[0115] 1) Determination of the image block of the peeling area: The identification of the peeling area is also based on the feature analysis of the image. An edge detection-based algorithm is used to determine the image block of the peeling area. For example, the Canny edge detection algorithm is used, which detects the edge by calculating the gradient of the image. Let the third UV topcoat performance detection image be I (x, y), and the Canny edge detection algorithm calculates the edge image E (x, y) of the image. Then, the peeling area is identified based on the edge features of the peeling area, such as the continuity and curvature of the edge. For example, the edge of the peeling area usually has a large curvature and discontinuity. By analyzing the edge image E (x, y), the area that meets these characteristics is identified as the peeling area image block. For example, in a 512*512 pixel image, an 80*80 pixel peeling area image block is identified.

[0116] (IV) Feature mapping and quantization encoding are performed on the image blocks in the peeling area to obtain the sixth quantization encoding data of each set performance analysis task cycle

[0117] 1) Feature mapping: For the feature mapping of the peeling area image block, the features such as the area A', edge length L and shape complexity C of the peeling area image block are considered. The shape complexity can be obtained by calculating the tortuosity of the edge, for example, by measuring it with the fractal dimension of the edge. For the area A', let its original range be [A_{min}', A_{max}'], for example [64, 6400], and use the linear mapping function f_{A'}(A')=frac{(A'-A_{min}')}{(A_{max}'-A_{min}')}times100 to map it to the range of [0, 100]. For the edge length L and shape complexity C, similar linear mapping functions are also used for mapping;

[0118] 2) Quantization coding: When quantizing and coding the features of the image blocks in the peeling area, a uniform quantization method is used. For example, the mapped area range [0, 100] is divided into 10 quantization intervals, each with a length of 10; the mapped edge length and shape complexity ranges are also divided into 10 quantization intervals. Encoding is performed according to the intervals in which the values ​​of the area, edge length, and shape complexity are mapped. For example, if the mapped value of the area is 25, the mapped value of the edge length is 35, and the mapped value of the shape complexity is 45, then the area is coded as 2, the edge length is coded as 3, and the shape complexity is coded as 4. These codes are combined to obtain the sixth quantization coding data for each set performance analysis task cycle.

[0119] (V) Fusing the fifth quantized coded data of each set performance analysis task cycle with the sixth quantized coded data of each set performance analysis task cycle to obtain a boundary texture visual description vector of each set performance analysis task cycle

[0120] 1) The significance of fusion processing: The purpose of fusing the fifth quantized coded data and the sixth quantized coded data is to combine the boundary texture information obtained from the peeling state and the flaking state to form a vector that fully describes the visual characteristics of the boundary texture. This can more accurately reflect the boundary texture of the UV topcoat within the set performance analysis task cycle;

[0121] 2) Fusion processing method: An exemplary fusion method is to arrange the fifth quantized coded data and the sixth quantized coded data in a certain order to form a vector. For example, if the fifth quantized coded data is [q_1, q_2, cdots, q_p] and the sixth quantized coded data is [r_1, r_2, cdots, r_q], then the boundary texture visual description vector can be expressed as [q_1, q_2, cdots, q_p, r_1, r_2, cdots, r_q]. The order here can be determined according to actual needs, for example, the codes related to the peeling state are arranged first, and then the codes related to the peeling state are arranged.

[0122] It can be seen that this embodiment, when setting the image description channel as the boundary texture visual description channel, comprehensively analyzes the boundary texture of the UV topcoat by respectively determining the relevant image features in the peeling state and the flaking state. In the peeling state analysis, the image blocks and thermal characteristics of the peeling area are calculated, and the image blocks of the flaking area are identified in the flaking state analysis, and the boundary texture information is deeply mined from different angles. The feature mapping and quantization encoding process standardizes the data and facilitates analysis. The quantization encoding data in the two states is integrated to form a boundary texture visual description vector, which integrates more boundary texture information, improves the accuracy and comprehensiveness of the UV topcoat boundary texture detection, and helps to more carefully capture the boundary texture changes of the topcoat within the set performance analysis task cycle, providing a reliable basis for accurately evaluating the performance of UV topcoat.

[0123] On the basis of the above content, the debugging step of the target performance evaluation decision algorithm includes: obtaining a number of second UV topcoat performance detection images within the set detection cycle upstream and downstream of the number of topcoat performance detection training images and a priori defect annotations on whether each topcoat performance detection training image has defects; for each topcoat performance detection training image, based on the set detection cycle, the set performance analysis task cycle and the set image description channel, the performance characterization vector mining of the number of second UV topcoat performance detection images is performed to obtain performance detection characterization vector set samples of the topcoat performance detection training images; through the initial performance evaluation decision algorithm, the performance evaluation decision is performed on the performance detection characterization vector set samples to obtain defect discrimination training information of the topcoat performance detection training images; if the defect discrimination training information does not match the priori defect annotation, the algorithm weight of the initial performance evaluation decision algorithm is debugged cyclically based on the algorithm training error of the initial performance evaluation decision algorithm until the initial performance evaluation decision algorithm is in a stable state; and the debugged initial performance evaluation decision algorithm is determined as the target performance evaluation decision algorithm.

[0124] In an embodiment of the present invention, an expanded introduction to the debugging steps of the target performance evaluation decision algorithm is as follows.

[0125] (I) Obtaining several second UV topcoat performance test images within the upstream and downstream of several topcoat performance test training images and setting the test cycle and prior defect annotations of whether each topcoat performance test training image has defects

[0126] 1) The significance of data acquisition: When constructing a target performance evaluation decision algorithm, obtaining accurate and comprehensive data is the basis. Several topcoat performance test training images are used as sample sets, and several second UV topcoat performance test images within the upstream and downstream set test cycles can reflect the performance of topcoats at different time stages. The prior defect annotation of whether each topcoat performance test training image has defects is an important basis for subsequent algorithm debugging, which is used to measure the accuracy of the algorithm output results;

[0127] 2) Data acquisition method: For the acquisition of the second UV topcoat performance test image, similar equipment and methods as those used to acquire the first UV topcoat performance test image can be used. For example, an industrial camera with a resolution of 1920*1080 pixels is used to acquire images under set lighting conditions (such as uniform diffuse light with a light intensity of 800 lux) at a certain time interval (such as acquiring an image every 30 minutes) within the upstream and downstream set test cycles (for example, the upstream test cycle is 0-12 hours, and the downstream test cycle is 12-24 hours).

[0128] The determination of prior defect annotations requires manual or known defect standard annotation. For example, the paint performance inspection training images with obvious defects such as scratches, uneven color, peeling, etc. are annotated as defective (marked as 1), while the images with intact surface and stable performance are annotated as non-defective (marked as 0).

[0129] (ii) For each topcoat performance detection training image, based on the set detection cycle, the set performance analysis task cycle and the set image description channel, the performance characterization vector mining is performed on several second UV topcoat performance detection images to obtain performance detection characterization vector set samples of the topcoat performance detection training images

[0130] 1) Image processing based on the set detection cycle and the set performance analysis task cycle: Similar to the previous processing of the UV topcoat performance detection image to be analyzed, the set detection cycle determines a total time range, and the set performance analysis task cycle is a subdivision unit within this total range. For example, the detection cycle is set to 24 hours, and the performance analysis task cycle is set to 6 hours. For the second UV topcoat performance detection image obtained within the 24-hour upstream and downstream set detection cycle of each topcoat performance detection training image, the image blocks are disassembled according to the 6-hour set performance analysis task cycle. For example, if an image is acquired every 30 minutes within 24 hours, a total of 48 second UV topcoat performance detection images are acquired. Then there are 12 images in each 6-hour set performance analysis task cycle, and these images are combined to form image blocks corresponding to each set performance analysis task cycle;

[0131] 2) Performance representation vector mining based on set image description channels: The set image description channels include mechanical deformation visual description channel, reflection characteristic visual description channel and boundary texture visual description channel, etc.

[0132] Mechanical deformation visual description channel: Determine the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel in the image block of each set performance analysis task cycle. For example, for particle deformation detection data, the particle structure is determined by an image segmentation algorithm, such as a threshold-based segmentation algorithm, and when the pixel grayscale value is greater than 120, it is determined to be a particle part. Then calculate the particle deformation, such as calculating the average coordinate offset of the particle contour coordinates overline{Deltax}=frac{1}{n}sum_{i=1}^{n}(x_i'-x_i) and overline{Deltay}=frac{1}{n}sum_{i=1}^{n}(y_i'-y_i) (for example, the particle contour coordinates at the initial moment are (x_i, y_i), and after the set performance analysis task cycle, they become (x_i', y_i')), as well as the contour area change rate DeltaS=frac{S'-S}{S} (S is the initial particle area, S' is the changed particle area).

[0133] For regional grayscale comparison detection data, the image is divided into 16×16 small areas, each area is 32×32 pixels, the average grayscale value of each area is calculated, and the average grayscale value change of the same area at different times is compared. Deltaoverline{G}_{ij}=overline{G}_{ij}^1-overline{G}_{ij}^0 (overline{G}_{ij}^0 is the average grayscale value at the initial moment, and overline{G}_{ij}^1 is the average grayscale value after the set performance analysis task cycle).

[0134] Then, feature mapping is performed on the regional grayscale comparison detection data, such as using a linear mapping function f(x)=ax+b (for example, mapping the range of Deltaoverline{G}_{ij} from [-50, 50] to [0, 100], then a=1, b=50, that is, f(Deltaoverline{G}_{ij})=Deltaoverline{G}_{ij}+50), and then quantization encoding is performed (such as uniform quantization, dividing [0, 100] into 10 quantization intervals, each interval length is 10, and encoding is performed according to the interval in which the value is located) The first quantized coded data is obtained; the particle deformation feature mining is performed on the particle deformation detection data (such as calculating the particle deformation direction angle theta=arctan(frac{overline{Deltay}}{overline{Deltax}}) and the standard deviation sigma_d of the distance change from each point on the particle contour to the centroid), and then quantized coding is performed (such as encoding by dividing the quantization interval according to the value range of different characteristic parameters) to obtain the second quantized coded data, and finally the first quantized coded data and the second quantized coded data are fused to obtain the mechanical deformation visual description vector.

[0135] Reflection characteristic visual description channel: When the image description channel is set as the reflection characteristic visual description channel (including the brightness distribution description channel and the spot state description channel), for the brightness distribution description channel, the gloss evaluation data is determined, such as using the grayscale weighted average method G_{eval}=frac{sum_{x,y}I(x,y)w(x,y)}{sum_{x,y}w(x,y)} (I(x,y) is the image pixel coordinate, w(x,y) is the weight function), and constructing the gloss area heat map (dividing the image into several small areas, calculating the average gloss of each area to construct a matrix). Then, feature mapping (such as linear mapping to a specific range) and quantization encoding (uniform quantization and encoding) are performed on the gloss evaluation data and the gloss area heat map to obtain the third quantization encoding data.

[0136] For the spot state description channel, determine the gloss evaluation data under the spot state task keywords (such as calculating the average brightness of the spot area G_{spot}=frac{1}{n}sum_{i=1}^{n}I(x_i, y_i) through the spot center brightness detection algorithm, where (x_i, y_i) are pixels in the spot area) and the gloss area heat map (divide the image into small areas, calculate the average brightness ratio of the spot in each area to construct a matrix), then perform feature mapping and quantization encoding to obtain the fourth quantization encoding data, and finally fuse the third quantization encoding data and the fourth quantization encoding data to obtain the reflectance characteristic visual description vector.

[0137] In terms of the boundary texture visual description channel: when the image description channel is set to the boundary texture visual description channel (including the peeling state visual description channel and the flaking state visual description channel), for the peeling state visual description channel, determine the peeling area image block (such as through an image segmentation algorithm based on texture differences, identifying the peeling area according to the grayscale co-occurrence matrix features) and the peeling area image thermal features (calculating the rate of change of pixel grayscale values ​​in the peeling area image block to construct a matrix), and then perform feature mapping (such as linear mapping of the area, perimeter, and thermal feature matrix elements of the peeling area image block) and quantization encoding (uniform quantization and encoding) to obtain the fifth quantization encoding data.

[0138] For the visual description channel of the peeling state, the image block of the peeling area is determined (such as identifying the peeling area through the Canny edge detection algorithm), and then the characteristics of the image block of the peeling area (such as area, edge length, shape complexity) are feature mapped (linearly mapped to a specific range) and quantized (uniformly quantized and encoded) to obtain the sixth quantized coded data, and finally the fifth quantized coded data and the sixth quantized coded data are fused to obtain the boundary texture visual description vector.

[0139] By combining the vectors obtained from the above different image description channels, we can obtain a set of performance detection representation vector samples for the topcoat performance detection training images.

[0140] (III) Perform performance evaluation and decision making on the performance detection representation vector set samples through the initial performance evaluation decision algorithm to obtain defect discrimination training information of the topcoat performance detection training image

[0141] 1) Selection of the initial performance evaluation decision algorithm: The initial performance evaluation decision algorithm can be the support vector machine (SVM) algorithm. The basic principle of the SVM algorithm is to find an optimal hyperplane in the feature space to separate data of different categories. For our problem, it is to separate the performance detection representation vector set samples corresponding to the topcoat performance detection training images with defects and without defects;

[0142] 2) Performance evaluation decision process: Input the performance detection characterization vector set sample into the SVM algorithm. Assume that the performance detection characterization vector set sample is vec{x}, and the SVM algorithm performs classification by calculating the value of the decision function f(vec{x})=vec{w}cdotvec{x}+b (where vec{w} is the weight vector and b is the bias term). If f(vec{x})>0, it is judged that there is no defect (output 0); if f(vec{x})leq0, it is judged that there is a defect (output 1). In this way, the defect discrimination training information of the topcoat performance detection training image is obtained.

[0143] (IV) If the defect discrimination training information does not match the prior defect annotation, the algorithm weights of the initial performance evaluation decision algorithm are debugged based on the algorithm training error of the initial performance evaluation decision algorithm until the initial performance evaluation decision algorithm is in a stable state

[0144] 1) Calculation of algorithm training error: The algorithm training error is defined as the ratio of the mismatch between the defect discrimination training information and the prior defect annotation. Assume that the defect discrimination training information is hat{y}, the prior defect annotation is y, and the number of training samples is N, then the algorithm training error is: E=frac{1}{N}sum_{i=1}^{N}|hat{y}_i-y_i|;

[0145] 2) Error-based weight debugging: When the algorithm training error is not 0, the weight vector vec{w} in the SVM algorithm needs to be adjusted. The gradient descent method can be used to adjust the weight. According to the loss function of SVM (such as hinge loss function L(vec{w}, b) = max(0, 1-y(vec{w}cdotvec{x}+b))), calculate the gradient nabla_{vec{w}}L of the loss function to the weight vector vec{w}, and then update the weight vector according to the formula vec{w}=vec{w}-alphanabla_{vec{w}}L (where alpha is the learning rate, which is a pre-set small positive number, such as 0.01). Repeat this process until the algorithm training error reaches a very small value (such as E<0.01) or the error no longer decreases significantly after multiple iterations (such as 100 times). At this time, the initial performance evaluation decision algorithm is considered to be in a stable state.

[0146] (V) Determine the debugged initial performance evaluation decision algorithm as the target performance evaluation decision algorithm

[0147] After the above debugging process, the initial performance evaluation decision algorithm that can accurately identify defects in training images (i.e., the algorithm training error meets the requirements) is determined as the target performance evaluation decision algorithm. This target performance evaluation decision algorithm can be used to make performance evaluation decisions on the UV topcoat performance detection images to be analyzed.

[0148] With this design, during the debugging process of the target performance evaluation decision algorithm, by obtaining comprehensive training image data and its prior defect annotations, an accurate basis is provided for algorithm debugging. Performance characterization vector mining is performed based on multiple settings to accurately describe the performance of the topcoat from multiple aspects to ensure the validity of the algorithm input data. The use of a suitable initial performance evaluation decision algorithm and weight debugging based on the algorithm training error enable the algorithm to accurately classify and improve the defect discrimination ability. The entire debugging process improves the accuracy and reliability of the target performance evaluation decision algorithm, and can more accurately analyze the UV topcoat performance detection image for performance evaluation decisions, thereby effectively evaluating the quality status of the UV topcoat.

[0149] In some other possible application scenarios, the several topcoat performance detection training images include several first training images and several second training images, the several first training images include several past UV topcoat performance detection images without topcoat performance defects, and the prior defect annotations of the first training images indicate that the first training images do not have topcoat performance defects; the several second training images include several past UV topcoat performance detection images with topcoat performance defects, and the prior defect annotations of the second training images indicate that the second training images have topcoat performance defects.

[0150] When constructing the target performance evaluation decision algorithm, several first training images play an important role as samples without topcoat performance defects. These images can provide characterization information of normal UV topcoat performance under various conditions, and provide a data basis for the algorithm to learn features under normal conditions. The prior defect annotation of the first training image clearly indicates that there is no topcoat performance defect. This annotation is based on a detailed analysis and evaluation of past UV topcoat performance test images. For example, in the past production process, a large number of UV topcoats were tested in many aspects, including hardness testing (such as using a hardness tester to measure, the hardness value is within a reasonable range, such as a pencil hardness of 2H-3H), wear resistance testing (tested by a wear tester, the wear amount is below the specified standard, such as a wear depth of less than 0.05mm), gloss testing (measured using a gloss meter, the gloss value meets the product requirements, such as a gloss between 80-90) and adhesion testing (using a cross-cut test, the adhesion level reaches level 0 or 1), etc. When these performance indicators all meet the corresponding quality standards, the corresponding UV topcoat performance inspection image is judged to be defect-free and marked as a priori defect annotation as defect-free (can be marked as 0).

[0151] For the first training image, performance characterization vector mining is also performed based on the set detection cycle, the set performance analysis task cycle and the set image description channel.

[0152] In terms of setting the detection cycle and setting the performance analysis task cycle: for example, the detection cycle is set to 24 hours, and the performance analysis task cycle is set to 6 hours. For each first training image, the previous UV topcoat performance detection images acquired within the 24-hour upstream and downstream set detection cycle (for example, an image is acquired once an hour, a total of 24 images are acquired), and the image blocks are disassembled according to the 6-hour set performance analysis task cycle. There are 6 images in each 6-hour task cycle, and these images are combined to form image blocks corresponding to each set performance analysis task cycle.

[0153] In terms of setting the image description channel: Mechanical deformation visual description channel: For each image block in the set performance analysis task cycle, determine the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel. In terms of particle deformation detection data, the particle structure is determined by an image segmentation algorithm (such as a segmentation algorithm based on region growth). For example, the grayscale value range of the image is [0, 255], the growth threshold is set to 120, and starting from the seed point (such as the pixel in the center of the image), the adjacent pixels with grayscale values ​​greater than 120 are merged into the particle area. Calculate the deformation of the particle. For example, for a particle, its center coordinates are (x_0, y_0) at the initial moment. After the set performance analysis task cycle, the center coordinates become (x_1, y_1). The coordinate offset is Deltax=x_1-x_0 and Deltay=y_1-y_0, and the area change rate of the particle is calculated at the same time. Assume that the initial particle area is S_0 and the area after change is S_1, then the area change rate is DeltaS=frac{S_1-S_0}{S_0}.

[0154] For regional grayscale comparison detection data, the image is divided into 10×10 small areas, each area is 51×51 pixels, the average grayscale value overline{G}_{ij} of each area is calculated, and the average grayscale value changes of the same area at different times are compared: Deltaoverline{G}_{ij}=overline{G}_{ij}^1-overline{G}_{ij}^0. Then, feature mapping is performed on the regional grayscale comparison detection data, for example, a logarithmic mapping function f(x)=ln(x+1) is used (the original grayscale value range is mapped to a new range for subsequent processing), and then quantization encoding is performed (such as using non-uniform quantization, dividing the mapped range into 8 quantization intervals according to the distribution characteristics of the data, and encoding is performed according to the interval where the value is located) to obtain the first quantization encoding data; particle deformation feature mining is performed on the particle deformation detection data (such as calculating the deformation direction angle theta=arctan(frac{Deltay}{Deltax}) of the particle and the standard deviation sigma_d of the distance change from each point on the particle contour to the centroid), and then quantization encoding is performed (such as dividing the quantization interval according to the value range of different feature parameters for encoding) to obtain the second quantization encoding data, and finally the first quantization encoding data and the second quantization encoding data are fused to obtain a mechanical deformation visual description vector.

[0155] Reflection characteristic visual description channel: When the image description channel is set as the reflection characteristic visual description channel (including the brightness distribution description channel and the spot state description channel), for the brightness distribution description channel, the gloss evaluation data is determined. Using the weighted average method, let the image be I (x, y), and the weight function be w (x, y) = frac {1} {d^2} (where d is the distance from the pixel (x, y) to the center of the image), then the gloss evaluation data G_{eval} = frac {sum_{x, y} I (x, y) w (x, y)} {sum_{x, y} w (x, y)}. When constructing the gloss area heat map, the image is divided into 8x8 small areas, and the average gloss of each area is calculated to construct a matrix. Then, the gloss evaluation data and the gloss area heat map are feature mapped (such as using an exponential mapping function to map the gloss evaluation data to a specific range) and quantized (using uniform quantization and encoding according to the mapped data range) to obtain the third quantized encoded data.

[0156] For the spot state description channel, determine the gloss evaluation data under the spot state task keyword (such as determining the spot center by finding the local maximum point in the image, and then calculating the average brightness in an area with a radius of 10 pixels and the spot center as the center of the circle as the gloss evaluation data under the spot state task keyword) and the gloss area heat map (divide the image into small areas, calculate the average brightness ratio of the spot in each area to construct a matrix), then perform feature mapping and quantization encoding to obtain the fourth quantization encoding data, and finally fuse the third quantization encoding data and the fourth quantization encoding data to obtain the reflectance characteristic visual description vector.

[0157] Boundary texture visual description channel: When the image description channel is set to the boundary texture visual description channel (including the peeling state visual description channel and the flaking state visual description channel), for the peeling state visual description channel, determine the peeling area image block (such as through an image segmentation algorithm based on texture energy, calculate the texture energy of the image, and determine it as a peeling area when the texture energy is lower than a certain threshold) and the peeling area image thermal features (calculate the rate of change of pixel grayscale values ​​in the peeling area image block to construct a matrix), and then perform feature mapping (such as linear mapping of the area, perimeter, and thermal feature matrix elements of the peeling area image block) and quantization encoding (uniform quantization and encoding) to obtain the fifth quantization encoding data.

[0158] For the peeling state visual description channel, the peeling area image block is determined (such as identifying the peeling area through an edge tracking-based algorithm, starting from the image edge, and tracking the boundary of the peeling area according to the direction and curvature change of the edge), and then the features of the peeling area image block (such as area, edge length, shape complexity) are feature mapped (linearly mapped to a specific range) and quantized (uniformly quantized and encoded) to obtain the sixth quantized encoded data, and finally the fifth quantized encoded data and the sixth quantized encoded data are fused to obtain the boundary texture visual description vector.

[0159] By combining the vectors obtained from the above different image description channels, we can obtain a sample set of performance detection representation vectors for the first training image.

[0160] Several second training images contain past UV topcoat performance inspection images with topcoat performance defects, which provide samples for learning defect features for the target performance evaluation decision algorithm. By analyzing the defect features in these images, the algorithm can better identify and distinguish defective UV topcoat performance inspection images.

[0161] The prior defect annotation of the second training image indicates that it has a topcoat performance defect. The determination of these defects is also based on multiple test results. For example, in hardness testing, if the pencil hardness is lower than 2H, it indicates insufficient hardness; in wear resistance testing, if the wear depth is greater than 0.05mm, it indicates poor wear resistance; in gloss testing, if the gloss value is not within the specified range (such as lower than 80 or higher than 90), it indicates that the gloss is abnormal; in adhesion testing, a cross-cut test result of level 2 or worse indicates poor adhesion. When the UV topcoat performance test image does not meet the standards in one or more of these performance indicators, it is determined to be defective and marked as defective (marked as 1) with the prior defect annotation.

[0162] Similar to the first training image, performance characterization vector mining is performed on the second training image based on a set detection cycle, a set performance analysis task cycle, and a set image description channel.

[0163] In terms of setting the detection cycle and setting the performance analysis task cycle: For example, if the detection cycle is set to 24 hours and the performance analysis task cycle is set to 6 hours, the previous UV topcoat performance detection images (obtained once per hour) acquired within the 24-hour upstream and downstream detection cycle of each second training image are decomposed into image blocks, and corresponding image blocks are formed within each 6-hour task cycle.

[0164] In terms of setting the image description channel: Mechanical deformation visual description channel: When determining the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel, due to defects, the particle deformation and grayscale comparison may have more obvious changes. For example, in terms of particle deformation detection data, in the case of insufficient hardness defects, the particles may be more easily squeezed and deformed. When calculating the deformation of particles, large coordinate offsets and area change rates may be found. For example, there is a particle with an initial coordinate of (x_0, y_0). After setting the performance analysis task cycle, the coordinates become (x_1, y_1). The calculation results are Deltax = x_1-x_0 = 5 pixels (which may be larger than normal), Deltay = y_1-y_0 = 3 pixels, the initial area is S_0 = 100 square pixels, and the area after the change is S_1 = 120 square pixels. The area change rate is DeltaS = frac{S_1-S_0}{S_0} = 0.2.

[0165] For regional grayscale comparison test data, the grayscale value changes will be different because defects may cause changes in the surface structure. Divide the image into 10×10 small areas, each area is 51×51 pixels, calculate the average grayscale value overline{G}_{ij} of each area, and compare the average grayscale value changes of the same area at different times: Deltaoverline{G}_{ij}=overline{G}_{ij}^1-overline{G}_{ij}^0. For example, it may be found that the grayscale value changes in some areas are much larger than normal. Then, the regional grayscale comparison detection data is feature mapped (such as using power function mapping f(x)=x^{0.5}) and quantized (such as using non-uniform quantization and dividing the data into 8 quantization intervals for encoding according to the distribution characteristics of the data) to obtain the first quantized coded data; the particle deformation detection data is subjected to particle deformation feature mining (such as calculating the particle deformation direction angle theta=arctan(frac{Deltay}{Deltax}) and the standard deviation sigma_d of the distance change from each point on the particle contour to the centroid), and then quantized (such as dividing the quantization interval for encoding according to the value range of different feature parameters) to obtain the second quantized coded data, and finally the first quantized coded data and the second quantized coded data are fused to obtain the mechanical deformation visual description vector.

[0166] Reflection characteristic visual description channel: For the reflection characteristic visual description channel (including brightness distribution description channel and spot state description channel), in the presence of defects, the gloss evaluation data and gloss area heat map will change significantly.

[0167] In the brightness distribution description channel, for gloss evaluation data, defects may affect the smoothness of the topcoat, resulting in uneven light reflection. When the weighted average method is used to calculate the gloss evaluation data, for example, the weight function is w(x, y) = frac{1}{d^2}, the calculated gloss evaluation data may be significantly different from the normal situation. When constructing the gloss area heat map, the image is divided into 8×8 small areas, and the average gloss of each area is calculated to construct a matrix. The gloss value of the defective area may be significantly lower or higher than the normal area. Then, feature mapping (such as using a logarithmic mapping function to map the gloss evaluation data to a specific range) and quantization encoding (using uniform quantization and encoding according to the mapped data range) are performed on the gloss evaluation data and the gloss area heat map to obtain the third quantization encoding data.

[0168] For the light spot state description channel, determine the gloss evaluation data under the light spot state task keyword (such as determining the light spot center by finding the local maximum point in the image, and then calculating the average brightness in the area with the light spot center as the center and a radius of 10 pixels as the gloss evaluation data under the light spot state task keyword) and the gloss area heat map (divide the image into small areas, calculate the average brightness proportion of the light spot in each area to build a matrix). Due to the existence of defects, the distribution and intensity of the light spot may change. Then perform feature mapping and quantization encoding to obtain the fourth quantization encoding data, and finally fuse the third quantization encoding data and the fourth quantization encoding data to obtain the reflection characteristic visual description vector.

[0169] Boundary texture visual description channel: In terms of the boundary texture visual description channel (including the peeling state visual description channel and the flaking state visual description channel), for the peeling state visual description channel, the peeling area may be more obvious when there is a defect. Determine the image block of the peeling area (such as through an image segmentation algorithm based on texture energy. Since the peeling causes the texture energy to change, when the texture energy is lower than a certain threshold, it is determined to be a peeling area) and the thermal characteristics of the image of the peeling area (calculate the gray value change rate of the pixels in the image block of the peeling area to build a matrix). The area and gray value change rate of the peeling area may be larger than normal. Then perform feature mapping (such as linear mapping of the area, perimeter, and thermal feature matrix elements of the peeling area image block) and quantization encoding (uniform quantization and encoding) to obtain the fifth quantization encoding data.

[0170] For the peeling state visual description channel, the peeling area image block is determined (such as identifying the peeling area through an edge tracking-based algorithm, as the peeling causes discontinuous and irregular edges), and then the features of the peeling area image block (such as area, edge length, shape complexity) are feature mapped (linearly mapped to a specific range) and quantized (uniformly quantized and encoded) to obtain the sixth quantized coded data, and finally the fifth quantized coded data and the sixth quantized coded data are fused to obtain the boundary texture visual description vector.

[0171] By combining the vectors obtained from the above different image description channels, we can obtain a sample set of performance detection representation vectors for the second training image.

[0172] In this way, by dividing the training images into the first training images without paint performance defects and the second training images with paint performance defects, the normal and abnormal states of UV paint are fully covered. When mining the performance characterization vector, starting from a variety of set image description channels, specific algorithms and numerical examples are used to accurately extract feature information under different states. This method provides rich and targeted learning samples for the target performance evaluation decision algorithm, which helps the algorithm to accurately learn the feature differences between normal and defective states, thereby improving the algorithm's judgment accuracy on whether there are defects in UV paint performance detection images, and enhancing the reliability and effectiveness of the entire UV paint performance detection system.

[0173] In an alternative embodiment, the method for obtaining the plurality of first training images includes: obtaining a first detection image queue including a plurality of past UV topcoat performance detection images without topcoat performance defects; determining from the first detection image queue a plurality of past UV topcoat performance detection images with a plurality of UV topcoat performance detection images within a set detection cycle of upstream and downstream to obtain a second detection image queue; arbitrarily screening the second detection image queue to obtain the plurality of first training images. Further, the method for obtaining the plurality of second training images includes: obtaining a third detection image queue including a plurality of past UV topcoat performance detection images with topcoat performance defects; grouping and summarizing the past UV topcoat performance detection images in the third detection image queue based on the types of defects in the past UV topcoat performance detection images to obtain a plurality of fourth detection image queues; arbitrarily screening each fourth detection image queue to obtain the plurality of second training images.

[0174] In the actual application process, in order to obtain the first training image, the first detection image queue must be obtained first. This queue contains several past UV topcoat performance detection images that do not have topcoat performance defects. These images may come from the database accumulated when a large number of UV topcoat products were previously tested for performance. For example, in a long-term UV topcoat production quality monitoring process, each batch of qualified UV topcoat products was subjected to a variety of performance tests and the corresponding test images were recorded. When these images were collected, specific equipment and parameters may have been used, such as using an industrial camera with a resolution of 1600*1200 pixels, and shooting the UV topcoat surface from different angles (such as 0°, 45°, 90°) under uniform lighting conditions with a light intensity of 600 lux to ensure that the state of the topcoat surface is fully and accurately recorded. These images constitute the first detection image queue.

[0175] From the first detection image queue, determine the past UV topcoat performance detection images with several UV topcoat performance detection images within the upstream and downstream set detection cycles, so as to obtain the second detection image queue. The set detection cycle here is a key parameter, which determines the time range of the selected image. For example, the detection cycle is set to 18 hours, the upstream detection cycle is 0-9 hours, and the downstream detection cycle is 9-18 hours. For each past UV topcoat performance detection image in the first detection image queue, it is necessary to check whether there are sufficient numbers (for example, at least 5) of related UV topcoat performance detection images within this 18-hour set detection cycle. These images may be acquired at certain time intervals (such as once every 3 hours) during this time period. Through such screening, past UV topcoat performance detection images with complete detection data within a specific time range can be obtained to form a second detection image queue.

[0176] The second detection image queue is randomly screened to obtain a number of first training images. This arbitrary screening can be to randomly select a certain number of images. For example, 20 past UV topcoat performance detection images are randomly selected from the second detection image queue as the first training images. The purpose of this is to further select representative image samples on the basis of meeting the time range and data integrity requirements for the subsequent training of the target performance evaluation decision algorithm.

[0177] First, obtain the third inspection image queue, which contains several past UV topcoat performance inspection images with topcoat performance defects. These defective images are also collected from past quality inspection records. For example, when spot-checking UV topcoat products, it is found that some products have problems such as insufficient hardness, poor wear resistance, abnormal gloss or poor adhesion. The UV topcoat performance inspection images corresponding to these defective products are recorded, and these images constitute the third inspection image queue.

[0178] Based on the types of defects in the past UV topcoat performance inspection images, the past UV topcoat performance inspection images in the third inspection image queue are grouped and summarized to obtain several fourth inspection image queues. For example, images with insufficient hardness defects are grouped together to form a fourth inspection image queue; images with poor wear resistance defects are grouped together to form another fourth inspection image queue; similarly, images with defects such as abnormal gloss and poor adhesion are also classified separately.

[0179] When determining the type of defect, it is necessary to rely on specific testing standards and algorithms. Taking hardness testing as an example, if the pencil hardness test method is used, when the pencil hardness is lower than the specified value (such as 2H), it is judged to be insufficient hardness. For wear resistance testing, if the wear amount exceeds the specified standard (such as wear depth greater than 0.05mm) tested by a wear tester, it is judged to be poor wear resistance. In gloss testing, if the gloss value measured by a gloss meter is not within the set range (such as lower than 80 or higher than 90), it is considered to be abnormal gloss. In terms of adhesion testing, through the cross-cut test, when the adhesion level is level 2 or worse, it is judged to be poor adhesion.

[0180] Through such grouping, each image in the fourth inspection image queue has the same type of defects, which helps to train the target performance evaluation decision algorithm more specifically, so that the algorithm can better identify different types of defect characteristics.

[0181] Each fourth detection image queue is randomly screened to obtain a number of second training images. For example, for each fourth detection image queue, 15 images can be randomly selected as second training images. This screening method can reduce the amount of data and improve training efficiency while ensuring that each defect type has enough representative images.

[0182] With this design, a specific method is used when acquiring the first training image and the second training image. For the first training image, multi-step screening is used to ensure that the selected image is a representative defect-free image within a specific detection cycle, providing an accurate sample for the target performance evaluation decision algorithm to learn normal topcoat performance. For the second training image, the grouping is summarized based on the defect type and then screened, so that the algorithm can learn for different defect types and enhance the ability to recognize various defects. This method of acquiring training images improves the accuracy and comprehensiveness of the training of the target performance evaluation decision algorithm, thereby improving the reliability and effectiveness of the algorithm in judging whether there are defects in the UV topcoat performance detection image.

[0183] In other preferred extensibility examples, the set image description channel also includes a UV curing process description channel, and the performance detection characterization vector to be analyzed also includes a UV curing process parameter vector. Based on this, under another optional design idea of ​​the above content, the initial performance evaluation decision algorithm includes a first decision tree branch and a second decision tree branch, the first decision tree branch includes a plurality of first fully connected modules, and the second decision tree branch includes a second fully connected module. The target performance evaluation decision algorithm performs a performance evaluation decision on the performance detection characterization vector set to be analyzed, and obtains a performance evaluation decision label of whether the UV topcoat performance detection image to be analyzed has defects, including: performing a performance evaluation decision on the performance detection characterization vector set to be analyzed through the plurality of first fully connected modules to obtain a plurality of performance evaluation decision features; weighting the plurality of performance evaluation decision features to obtain a performance evaluation weighted decision feature; performing a performance evaluation decision on the performance evaluation weighted decision feature through the second fully connected module to obtain the performance evaluation decision label.

[0184] When the image description channel is set to include a UV curing process description channel, the technical solution of the related performance evaluation decision algorithm is introduced as follows.

[0185] (I) Setting the UV curing process description channel in the image description channel and the UV curing process parameter vector in the performance detection characterization vector to be analyzed

[0186] 1) The significance of UV curing process description channel: When testing and analyzing the performance of UV topcoat, the introduction of UV curing process description channel provides a new dimension to comprehensively evaluate the performance of topcoat. The UV curing process has a crucial influence on the final performance of UV topcoat, including hardness, wear resistance, gloss and adhesion. Different curing process parameters may lead to significant differences in the microstructure and macro properties of the topcoat;

[0187] 2) Determination of UV curing process parameter vector: The UV curing process parameter vector is used as part of the performance detection characterization vector to be analyzed to quantify the characteristics related to the UV curing process. These parameters may include curing time, curing temperature, UV light intensity, etc. For example, the curing time can vary between 1-10 seconds, the curing temperature between 40-80°C, and the UV light intensity between 100-500mW / cm². For example, for a specific UV topcoat sample, its curing time is 5 seconds, the curing temperature is 60°C, and the UV light intensity is 300mW / cm². These parameters can be normalized. For example, for the curing time, the formula is: t_{norm}=frac{t-t_{min}}{t_{max}-t_{min}}, where t is the actual curing time, t_{min} is the minimum curing time (1 second), and t_{max} is the maximum curing time (10 seconds). The curing time of the sample is normalized to: t_{norm}=frac{5-1}{10-1}=frac{4}{9}. Similarly, the curing temperature and UV light intensity are normalized to obtain the corresponding normalized values, and these normalized parameters are combined into a UV curing process parameter vector, such as [frac{4}{9}, frac{20}{40}, frac{200}{400}].

[0188] (II) The first decision tree branch and the second decision tree branch in the initial performance evaluation decision algorithm

[0189] 1) First decision tree branch - several first fully connected modules: Several first fully connected modules in the first decision tree branch play an important role in performance evaluation decisions. A fully connected module is a neural network structure that connects each element in the input vector (here is the performance test characterization vector set to be analyzed) with each neuron in the next layer. For example, the performance test characterization vector set to be analyzed contains multiple sub-vectors, such as the hardness performance test characterization vector set, the wear resistance test characterization vector set, the gloss test characterization vector set, the adhesion test characterization vector set, and the UV curing process parameter vector mentioned above. These vectors are combined into an input vector vec{x} and input into the first first fully connected module of the first decision tree branch. Assume that the first first fully connected module has n_1 input neurons and m_1 output neurons, its weight matrix is ​​W_1, and the bias vector is vec{b}_1. For the input vector vec{x}, after calculation by the first first fully connected module, the output vector vec{y}_1=f(W_1vec{x}+vec{b}_1) is obtained, where f is the activation function, for example, the ReLU function f(x)=max(0,x) can be used. This output vector vec{y}_1 is a performance evaluation decision feature. Similarly, vec{y}_1 is used as the input of the next first fully connected module, and the calculation continues to obtain more performance evaluation decision features. For example, after calculation by k first fully connected modules, k performance evaluation decision features vec{y}_1, vec{y}_2, cdots, vec{y}_k are obtained;

[0190] 2) Weighting of performance evaluation decision features - obtaining performance evaluation weighted decision features: Weighting the obtained several performance evaluation decision features is to comprehensively consider the importance of each feature. Suppose the weighted vector is vec{w}=[w_1, w_2, cdots, w_k], where w_i represents the weight of the i-th performance evaluation decision feature. The performance evaluation weighted decision feature vec{z} can be calculated by the formula vec{z}=sum_{i=1}^{k}w_ivec{y}_i. These weights can be determined through prior training or based on experience. For example, if it is found in previous experiments and data analysis that a certain performance evaluation decision feature (such as a feature related to hardness) is more critical to the final defect judgment, a larger weight can be given. For example, w_1=0.3, w_2=0.2, w_3=0.1 (this is just an example), then the performance evaluation weighted decision feature vec{z} is calculated according to the above formula;

[0191] 3) Second decision tree branch - second fully connected module: The second fully connected module in the second decision tree branch receives the performance evaluation weighted decision feature vec{z} as input. Assume that the second fully connected module has n_2 input neurons and 1 output neuron, its weight matrix is ​​W_2, and its bias vector is vec{b}_2. After calculation by the second fully connected module, the output value y=f(W_2vec{z}+vec{b}_2) is obtained. This output value y is the final performance evaluation decision label. For example, if y>0.5, it is determined that there are defects in the UV topcoat performance detection image to be analyzed (marked as 1); if yleq0.5, it is determined that there are no defects (marked as 0).

[0192] In this way, the UV curing process description channel and the corresponding UV curing process parameter vector are added to the set image description channel, which provides more information for the UV topcoat performance analysis from the perspective of curing process, and helps to evaluate the topcoat performance more comprehensively. The initial performance evaluation decision algorithm structure including the first decision tree branch (multiple first fully connected modules) and the second decision tree branch (second fully connected module) is adopted. The performance evaluation decision features are gradually extracted and integrated through multiple fully connected modules, and weighted processing is performed to finally obtain the performance evaluation decision label. This structure can more effectively utilize the information of various performance detection characterization vectors, improve the accuracy and reliability of performance evaluation decisions, and enhance the ability to judge whether there are defects in the performance of UV topcoat.

[0193] Further, Figure 2 FIG. 2 is a schematic diagram of a UV topcoat performance detection and analysis device 200 provided in an embodiment of the present invention. Figure 2 The UV topcoat performance detection and analysis device 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method in the embodiment of the present invention. Figure 2 As shown, the UV topcoat performance detection and analysis device 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention. The memory 230 may be a separate device independent of the processor 210, or may be integrated in the processor 210. Optionally, as Figure 2As shown, the UV topcoat performance detection and analysis device 200 may also include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices. Optionally, the UV topcoat performance detection and analysis device 200 may implement the corresponding processes corresponding to the storage engine or the components (such as processing modules) in the storage engine or the device deployed with the storage engine in each method of the embodiment of the present invention, and for the sake of brevity, it will not be repeated here. It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capabilities. It can be understood that the memory in the embodiment of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. It should be noted that the memory of the system and method described herein is intended to include, but is not limited to, a memory of a suitable type.

[0194] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and the program or instruction implements the steps of the above method when executed by a processor.

[0195] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the embodiments of the present invention are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present invention and the scope of protection of the embodiments of the present invention, all of which are within the protection of the embodiments of the present invention.

Claims

1. A UV topcoat performance detection and analysis method based on image data processing, characterized in that: The method is applied to a UV topcoat performance detection and analysis device, and the method comprises: obtaining a plurality of first UV topcoat performance detection images within a detection cycle set upstream and downstream of the UV topcoat performance detection image to be analyzed; performing performance characterization vector mining on the plurality of first UV topcoat performance detection images based on the set detection cycle, the set performance analysis task cycle and the set image description channel, and obtaining a set of performance detection characterization vectors to be analyzed of the UV topcoat performance detection image to be analyzed; the set performance analysis task cycle is less than the set detection cycle; the set of performance detection characterization vectors to be analyzed comprises a hardness performance detection characterization vector set, a wear resistance detection characterization vector set, a gloss detection characterization vector set and an adhesion detection characterization vector set; wherein, within the set detection cycle, the plurality of first UV topcoat performance detection images are image block disassembled according to the set performance analysis task cycle, and a third UV topcoat performance detection image of each set performance analysis task cycle within the set detection cycle is obtained; performing performance characterization vector mining on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel, and obtaining a set of performance detection characterization vectors to be analyzed of each set performance analysis task cycle; and performing performance characterization vector mining on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set detection cycle. The performance detection characterization vectors to be analyzed of several set performance analysis task cycles are integrated and interacted to obtain the performance detection characterization vector set to be analyzed; the set image description channel includes one or more of a mechanical deformation visual description channel, a reflection characteristic visual description channel and a boundary texture visual description channel; when the set image description channel includes the mechanical deformation visual description channel, the performance detection characterization vector to be analyzed includes a mechanical deformation visual description vector; when the set image description channel includes the reflection characteristic visual description channel, the performance detection characterization vector to be analyzed includes a reflection characteristic visual description vector; when the set image description channel includes the boundary texture visual description channel, the performance detection characterization vector to be analyzed includes a boundary texture visual description vector; a performance evaluation decision is made on the performance detection characterization vector set to be analyzed by a target performance evaluation decision algorithm to obtain a performance evaluation decision label of whether the UV topcoat performance detection image to be analyzed has defects; the target performance evaluation decision algorithm is obtained by debugging an initial performance evaluation decision algorithm based on several second UV topcoat performance detection images within the upstream and downstream set detection cycles of several topcoat performance detection training images and a priori defect annotations of whether each topcoat performance detection training image has defects.

2. The method according to claim 1, characterized in that When the set image description channel includes the mechanical deformation visual description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the particle deformation detection data and regional grayscale comparison detection data corresponding to the mechanical deformation visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; performing feature mapping and quantization encoding on the regional grayscale comparison detection data to obtain the first quantization encoding data for each set performance analysis task cycle; performing particle deformation feature mining and quantization encoding on the particle deformation detection data to obtain the second quantization encoding data for each set performance analysis task cycle; performing fusion processing on the first quantization encoding data of each set performance analysis task cycle and the second quantization encoding data of each set performance analysis task cycle to obtain the mechanical deformation visual description vector for each set performance analysis task cycle.

3. The method according to claim 1, characterized in that When the set image description channel includes the reflective characteristic visual description channel, and the reflective characteristic visual description channel includes a brightness distribution description channel and a spot state description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the gloss evaluation data and gloss area heat map under the brightness distribution task keyword corresponding to the brightness distribution description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; feature mapping and quantification of the gloss evaluation data and gloss area heat map under the brightness distribution task keyword Encoding to obtain the third quantized coded data of each set performance analysis task cycle; determining the gloss evaluation data and the gloss area heat map under the spot state task keywords corresponding to the spot state description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; feature mapping and quantization encoding of the gloss evaluation data and the gloss area heat map under the spot state task keywords to obtain the fourth quantized coded data of each set performance analysis task cycle; fusing the third quantized coded data of each set performance analysis task cycle and the fourth quantized coded data of each set performance analysis task cycle to obtain the reflective characteristic visual description vector of each set performance analysis task cycle.

4. The method according to claim 1, characterized in that When the set image description channel includes the boundary texture visual description channel, and the boundary texture visual description channel includes a peeling state visual description channel and a flaking state visual description channel, the performance characterization vector mining is performed on the third UV topcoat performance detection image of each set performance analysis task cycle based on the set image description channel to obtain the performance detection characterization vector to be analyzed for each set performance analysis task cycle, including: determining the peeling area image block and the peeling area image thermal features corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle; The thermal characteristics are feature mapped and quantized to obtain the fifth quantized coded data of each set performance analysis task cycle; the peeling area image block corresponding to the peeling state visual description channel in the third UV topcoat performance detection image of each set performance analysis task cycle is determined; the peeling area image block is feature mapped and quantized to obtain the sixth quantized coded data of each set performance analysis task cycle; the fifth quantized coded data of each set performance analysis task cycle and the sixth quantized coded data of each set performance analysis task cycle are fused to obtain the boundary texture visual description vector of each set performance analysis task cycle.

5. The method according to claim 1, characterized in that The debugging step of the target performance evaluation decision algorithm includes: obtaining a plurality of second UV topcoat performance detection images within a set detection cycle upstream and downstream of the plurality of topcoat performance detection training images and a priori defect annotations of whether each of the topcoat performance detection training images has defects; for each of the topcoat performance detection training images, based on the set detection cycle, the set performance analysis task cycle and the set image description channel, performing performance characterization vector mining on the plurality of second UV topcoat performance detection images to obtain performance detection characterization vector set samples of the topcoat performance detection training images; performing performance evaluation decisions on the performance detection characterization vector set samples through the initial performance evaluation decision algorithm to obtain defect discrimination training information of the topcoat performance detection training images; if the defect discrimination training information does not match the priori defect annotations, cyclically debugging the algorithm weights of the initial performance evaluation decision algorithm based on the algorithm training error of the initial performance evaluation decision algorithm until the initial performance evaluation decision algorithm is in a stable state; and determining the debugged initial performance evaluation decision algorithm as the target performance evaluation decision algorithm.

6. The method according to claim 1, characterized in that The plurality of topcoat performance detection training images include a plurality of first training images and a plurality of second training images, the plurality of first training images include a plurality of past UV topcoat performance detection images without topcoat performance defects, and the prior defect annotations of the first training images indicate that the first training images do not have topcoat performance defects; the plurality of second training images include a plurality of past UV topcoat performance detection images with topcoat performance defects, and the prior defect annotations of the second training images indicate that the second training images have topcoat performance defects.

7. The method according to claim 6, characterized in that The method for obtaining the plurality of first training images includes: obtaining a first detection image queue including a plurality of past UV topcoat performance detection images without topcoat performance defects; determining from the first detection image queue a plurality of past UV topcoat performance detection images having a plurality of UV topcoat performance detection images within a set detection cycle of upstream and downstream, to obtain a second detection image queue; arbitrarily screening the second detection image queue to obtain the plurality of first training images; the method for obtaining the plurality of second training images includes: obtaining a third detection image queue including a plurality of past UV topcoat performance detection images with topcoat performance defects; grouping and summarizing the past UV topcoat performance detection images in the third detection image queue based on the types of defects in the past UV topcoat performance detection images, to obtain a plurality of fourth detection image queues; and arbitrarily screening each fourth detection image queue to obtain the plurality of second training images.

8. A UV topcoat performance detection and analysis device, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 7.

Images