AI coating edge catching processing method and system based on full connection network classification model

Abstract

The application provides an AI coating edge grabbing processing method and system based on a full connection network classification model, combines a traditional algorithm and an advanced machine learning algorithm, improves the accuracy of edge grabbing identification, and improves the robustness and stability of identification; compared with a traditional algorithm, the application eliminates the influence of subjective factors caused by individual differences and physical conditions of workers on defect judgment, the influence of uneven coating edges and variable types on the accuracy of identification results, and the difficulty in identifying virtual edges, solves the problem of poor key point detection effect of an artificial intelligence key point algorithm on an image edge and the problem of over-detection of an image edge after filling, and effectively solves the problem of low accuracy of coating edge grabbing identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an AI coating edge grabbing processing method and system of a full connection network classification model. BACKGROUND

[0002] When the sol is pasted on the coating area and the at9 area of the coating, three kinds of edges will be formed, which are virtual edge, fusion edge and solid edge. The virtual edge is the junction of the lug rubber and the at9 area; the fusion edge is the boundary line between the coating area covered by the sol and the at9 area; and the solid edge is the boundary line between the sol and the coating area. The purpose of edge grabbing is to grab the accurate positions of the virtual edge, the fusion edge and the solid edge by using traditional algorithms or AI algorithms, so as to measure the film size.

[0003] In the prior art, the traditional edge grabbing algorithm needs to adjust various parameters for a specific edge, has high operation cost, and has the risk of edge grabbing failure caused by human operation errors. When the coating edge is uneven and the type is variable, the desired position cannot be grabbed, and the fluctuation is large.

[0004] The technical defects of the prior art are as follows:

[0005] 1. The traditional edge grabbing algorithm needs to adjust various parameters for a specific edge, has high operation cost, and has the risk of edge grabbing failure caused by human operation errors.

[0006] 2. The traditional edge grabbing algorithm cannot grab the desired position when the coating edge is uneven and the type is variable, and the fluctuation is large.

[0007] 3. The traditional edge grabbing algorithm is difficult to grab the virtual edge because of small pixel difference, and it is difficult to flexibly select the virtual edge, the fusion edge or the solid edge to be grabbed.

[0008] 4. The traditional edge grabbing algorithm has large fluctuation and poor robustness when there are special conditions such as foil exposure or conversion between the edges that can be grabbed.

[0009] 5. The AI key point algorithm has poor key point detection effect on the edges of the image.

[0010] Therefore, it is necessary to provide a new coating edge grabbing processing method and system with higher accuracy, stability and robustness. SUMMARY

[0011] Therefore, in order to solve the above problems in the prior art, the present application provides an AI coating edge grabbing processing method and system based on a full connection network classification model.

[0012] The present application solves the above problems by the following technical means:

[0013] The first aspect of the present application provides an AI coating edge grabbing processing method based on a full connection network classification model, characterized by comprising the following steps:

[0014] Step S100, acquiring a coating edge image and extracting corresponding key point features;

[0015] Step S200, analyzing the key point features using convolution and full connection operations, and determining the longitudinal coordinate, transverse coordinate, and visibility of the key points through a classification head;

[0016] Step S300, constructing a full connection network classification model and training until the loss function converges, inputting the longitudinal coordinate, transverse coordinate, and visibility of the key points into the full connection network classification model, and extracting the key point coordinates of the virtual edge, fusion edge, and real edge at the edge of the coating edge image;

[0017] Step S400, inputting the longitudinal coordinate, transverse coordinate, and visibility of the key points of the coating edge image into the trained full connection network classification model for detection and obtaining the key point coordinates of the virtual edge, fusion edge, and real edge at the edge of the coating edge image, and determining the positions of the virtual edge, fusion edge, and real edge according to the key point coordinates.

[0018] Preferably, the acquiring of the coating edge image and the extraction of the corresponding key point features comprise the following steps:

[0019] Acquiring a coating original image and performing preprocessing to convert the picture into a unified format, the preprocessing including one or more of the following ways: scaling, cropping, grayscale enhancement, contrast enhancement;

[0020] Determining the first edge position of the preprocessed coating original image through an edge extraction algorithm, and cropping the coating original image with the first edge position as the center to generate a coating edge image;

[0021] Data labeling is performed on the coating edge image, and a feature extraction network with a pre-trained model is used to extract the key point features of the coating edge image.

[0022] Preferably, the analysis of the key point features using convolution and full connection operations and the determination of the longitudinal coordinate, transverse coordinate, and visibility of the key points through the classification head comprise:

[0023] The classification head comprises a longitudinal coordinate classification head, a transverse coordinate classification head, and a visibility classification head;

[0024] The longitudinal coordinate of the key point is determined through the longitudinal coordinate classification head, the transverse coordinate of the key point is determined through the transverse coordinate classification head, and the visibility of the key point is determined through the visibility classification head.

[0025] Preferably, the analysis of the key point features using convolution and full connection operations and the determination of the longitudinal coordinate, transverse coordinate, and visibility of the key points through the classification head further comprise:

[0026] The ordinate classification head transforms the image features into N one-dimensional vectors through one-dimensional convolution and full connection operation, representing the ordinate of the key points;

[0027] The abscissa classification head transforms the image features into N one-dimensional vectors through one-dimensional convolution and full connection operation, representing the abscissa of the key points;

[0028] The visibility classification head transforms the image features into a one-dimensional vector with a length of N through two-dimensional convolution and full connection operation, representing the visibility of the key points.

[0029] Preferably, the determination of the ordinate of each key point through the ordinate classification head, the determination of the abscissa of each key point through the abscissa classification head, and the determination of the visibility of each key point through the visibility classification head further comprise:

[0030] The ordinate classification head first compresses the extracted key point features through two-dimensional convolution and ReLU activation function; then flattens the compressed features in the abscissa direction, performs full connection operation and ReLU activation function, and converts the feature dimension to a times of the height of the coating edge image; finally, performs scale-invariant one-dimensional convolution operation in the abscissa direction to obtain the output result of the ordinate classification head of the key point.

[0031] The abscissa classification head first compresses the extracted key point features through two-dimensional convolution and ReLU activation function; then flattens the compressed features in the ordinate direction, performs full connection operation and ReLU activation function, and converts the feature dimension to a times of the width of the original image; finally, performs scale-invariant one-dimensional convolution operation in the ordinate direction to obtain the output result of the abscissa classification head of the key point.

[0032] The visibility classification head first compresses the extracted key point features through two-dimensional convolution twice, then performs global pooling, flattening, and full connection operation to obtain the output result of the visibility classification head of the key point.

[0033] Preferably, the obtaining of the coating edge image and the extraction of the corresponding key point features comprise:

[0034] The coating edge image is obtained, and pixels of a specific color are gradually filled outward on the edge of the image, and the color of each circle of pixels is the same, and the difference in chroma and saturation between adjacent circle layers is set within a specific range;

[0035] The key point labels located on the edge of the image in the image labeling data are moved to the outermost circle of the filled pixel area.

[0036] As preferred, the full connection network classification model needs to move the predicted key points to the edge of the image according to different key point types when performing inference;

[0037] The key points extracted by the full connection network classification model include two key point coordinates of the virtual edge, the real edge and the fusion edge at the edge position of the image, and the positions of the virtual edge, the fusion edge and the real edge are determined according to the principle of determining a straight line according to two points.

[0038] The second aspect of the application provides an AI coating edge catching processing system based on a full connection network classification model, comprising:

[0039] A processing module is configured to acquire a coating edge image and extract corresponding key point features.

[0040] An analysis module is configured to analyze the key point features by convolution and full connection operation, and determine the longitudinal coordinate, the horizontal coordinate and the visibility of the key points through a classification head.

[0041] A training module is configured to construct a full connection network classification model and train the model until the loss function converges, input the longitudinal coordinate, the horizontal coordinate and the visibility of the key points into the full connection network classification model, and extract the key point coordinates of the virtual edge, the fusion edge and the real edge at the edge of the coating edge image.

[0042] A detection module is configured to input the longitudinal coordinate, the horizontal coordinate and the visibility of the key points of the coating edge image into the trained full connection network classification model to detect and acquire the key point coordinates of the virtual edge, the fusion edge and the real edge at the edge of the coating edge image, and determine the positions of the virtual edge, the fusion edge and the real edge according to the key point coordinates.

[0043] The third aspect of the application provides a computer device, comprising a processor and a memory, the memory is configured to store executable instructions, the instructions are configured to control the processor to execute the method of any one of the first aspect of the application.

[0044] The fourth aspect of the application provides a computer readable storage medium, which stores a computer program, the computer program is executed by a processor to implement the method of any one of the first aspect of the application.

[0045] Compared with the prior art, the application has at least the following beneficial effects:

[0046] The application provides an AI coating edge grabbing processing method and system based on a full connection network classification model, which combines a traditional algorithm and an advanced machine learning algorithm, improves the accuracy of edge grabbing identification, and improves the robustness and stability of identification; compared with a traditional algorithm, the application eliminates the influence of subjective factors caused by individual differences and physical conditions of workers on defect judgment, the influence of uneven coating edges and various types on the accuracy of identification results, and the difficulty in identifying virtual edges, solves the problem of poor key point detection effect of an artificial intelligence key point algorithm on an image edge and the problem of over-detection of an image edge after filling, and effectively solves the problem of low accuracy of coating edge grabbing identification. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed to be used in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative effort on the basis of these drawings.

[0048] Figure 1 is a flowchart of a coating key point identification model provided by the application;

[0049] Figure 2 is a schematic diagram of an AI key point model classification head with full connection provided by the application;

[0050] Figure 3 is a schematic diagram of pixel filling of an image provided by the application;

[0051] Figure 4 is a schematic diagram of image features of an input model provided by the application;

[0052] Figure 5 is an output schematic diagram of a transverse coordinate classification head without adding a full connection network provided by the application;

[0053] Figure 6 is an output schematic diagram of a transverse coordinate classification head with added full connection network provided by the application;

[0054] Figure 7 is a flowchart of an AI coating edge grabbing processing method based on a full connection network classification model provided by the application;

[0055] Figure 8 is a structural schematic diagram of an AI coating edge grabbing processing system based on a full connection network classification model provided by the application;

[0056] Figure 9The application provides a computer equipment structure schematic diagram of AI coating edge catching treatment based on a full connection network classification model. DETAILED DESCRIPTION

[0057] In order to make the above objectives, characteristics and advantages of the present application more apparent, clear and understandable, the technical solutions of the present application will be described in detail below with reference to the drawings and specific embodiments. It should be pointed out that the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0058] In the present application, the relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0059] Technical terms involved in the present application:

[0060] ReLU activation: ReLU (Rectified Linear Unit) is a commonly used activation function in deep learning models. The definition of ReLU function is: when the input value is greater than 0, the output value is equal to the input value; when the input value is less than or equal to 0, the output value is equal to 0. The image of ReLU function is like a straight line, which has the characteristics of simplicity, speed and effectiveness. In deep learning models, ReLU function is usually used as the activation function of hidden layer, which can accelerate model training and improve model performance. The advantage of ReLU function lies in its nonlinearity, which can simulate complex input-output relationships; at the same time, it has sparsity, which can reduce the number of model parameters, thereby improving the efficiency and generalization ability of the model.

[0061] Feature flattening: Feature flattening is the process of converting multi-dimensional data into one-dimensional data, commonly used in feature extraction and processing in deep learning models. In deep learning models, input data is usually a multi-dimensional array composed of multiple features, such as image data, which is usually a two-dimensional array composed of multiple pixel values. In the feature flattening process, the multi-dimensional array is converted into a one-dimensional array, that is, each feature is taken as a separate element, and then all feature elements are combined into a one-dimensional array. Feature flattening can convert multi-dimensional data into one-dimensional data, so that deep learning models can process and process these data. For example, in a convolutional neural network, input image data usually needs to be flattened before being processed by a convolutional layer. In a recurrent neural network, input sequence data also needs to be flattened before being processed by an RNN layer.

[0062] Key point detection: Key point detection is a computer vision technique used to detect and locate key points with specific features in an image. These key points are usually feature points with significant shape, edge or texture in the image. Key point detection can automatically find these feature points in the image, providing a simple and effective method to identify and analyze objects and scenes in the image. In industrial image processing, key point detection can be used to detect defects, wear and tear or other abnormalities on products or parts, as well as to detect the position and movement of objects or equipment on the production line.

[0063] Embodiment one

[0064] Reference Figure 1 With Figure 2 The model structure diagram shown in the figure, and Figure 7 The flowchart, the embodiment of the application provides an AI coating edge handling method based on a full connection network classification model, characterized in that it comprises the following steps:

[0065] Step S100, acquiring the coating edge image and extracting the corresponding key point feature.

[0066] As a preferred, the acquisition of the coating edge image and the extraction of the corresponding key point feature, comprising the following steps:

[0067] Acquire the coating original image and perform preprocessing to convert the picture to a unified format, the preprocessing includes one or more of the following ways: scaling, cropping, grayscale enhancement, contrast enhancement;

[0068] Determine the first edge position of the preprocessed coating original image by edge extraction algorithm, crop the coating original image with the first edge position as the center, and generate the coating edge image;

[0069] Data labeling is performed on the coating edge image, and the key point feature of the coating edge image is extracted using a feature extraction network with a pre-trained model.

[0070] wherein, Figure 1 and Figure 2 H represents the image height, W represents the image width, N represents the number of types of key points, and a represents a hyperparameter for controlling sub-pixel accuracy.

[0071] Through the above steps, the original image photographed by the coating factory on-site camera is preprocessed, and the approximate position of the edge is obtained by an edge extraction algorithm on the processed image, and the edge extraction algorithm includes one or more of first-order derivative or second-order derivative.

[0072] The first-order derivative and the second-order derivative here refer to the concept in image processing. The first-order derivative represents the difference between each pixel in the image and its surrounding pixels, reflecting the trend of gray level change in the image. The second-order derivative represents the speed of gray level change in the image, which can be used to detect sharp changes or inflection points in the image. In the first-order or second-order derivative edge extraction algorithm, the position information of the edge in the image can be obtained by calculating the first-order or second-order derivative of the image.

[0073] According to the load and computing power of the model, the picture of a specific size around the edge is cropped and labeled for training the artificial intelligence key point detection algorithm model. In one possible implementation, considering the model load, a large number of small pictures of 640x640 around the approximate position of the edge can be cropped and labeled.

[0074] Reference Figure 3 As preferred, the acquiring of the coating edge image and the extraction of the corresponding key point features include:

[0075] The coating edge image is acquired, and a specific color of pixels is gradually filled outward on the edge of the image, and the color of each circle of pixels is the same, and the color difference and saturation difference between adjacent circles are set within a specific range;

[0076] The key point label located on the edge of the image in the image labeling data is moved to the outermost circle of the filled pixel area.

[0077] Figure 3 The two images on the left are image examples before filling, and the two images on the right are image examples after filling; the two images on the top are partial enlarged views of the two images on the bottom. By performing specific pixel filling on the coating edge image, the filled pixels act as the edges of the new image, which enhances the detection effect of the artificial intelligence key point detection algorithm model on the key points at the edge position of the image before filling; at the same time, by setting the color difference and saturation difference between adjacent circles of filled pixels within a specific range, the feature discrimination degree between the filled pixels and the edge position of the image before filling is enhanced, which facilitates the model to learn the key point features of the virtual edge, the fusion edge and the real edge at the edge position of the image.

[0078] Reference Figure 3 The upper right corner figure, after filling the pixels, the key point label in the image edge in the image annotation data is moved to the outermost circle of the filled pixel area, and then input to the subsequent step for key point feature extraction;

[0079] In step S200, the key point features are analyzed by convolution and full connection operation, and the longitudinal coordinate, horizontal coordinate and visibility of the key point are determined by the classification head.

[0080] Convolution: Convolution is a mathematical operation used in image and signal processing to extract features and patterns. In the convolution process, a small matrix called a convolution kernel or filter is applied to each position of the input image, resulting in a new output image where each pixel value is the weighted average of the overlapping area of the input image with the convolution kernel. Convolution can be used to detect edges, corners, textures and other features in images, and can also be used for image classification, object detection, image segmentation and other tasks. In computer vision, Convolutional Neural Networks (CNN) is a commonly used deep learning model that uses convolution layers to extract image features and combines and abstracts features between multiple convolution layers to achieve efficient classification and recognition of images.

[0081] As a preferred, the key point features are analyzed by convolution and full connection operation, and the longitudinal coordinate, horizontal coordinate and visibility of the key point are determined by the classification head, including:

[0082] The classification head includes a longitudinal coordinate classification head, a horizontal coordinate classification head and a visibility classification head;

[0083] The longitudinal coordinate of each key point is determined by the longitudinal coordinate classification head, the horizontal coordinate of each key point is determined by the horizontal coordinate classification head, and the visibility of each key point is determined by the visibility classification head.

[0084] Further, the key point features are analyzed by convolution and full connection operation, and the longitudinal coordinate, horizontal coordinate and visibility of the key point are determined by the classification head, further including:

[0085] The longitudinal coordinate classification head transforms the image features into N one-dimensional vectors by one-dimensional convolution and full connection operation, representing the longitudinal coordinate of each key point;

[0086] The horizontal coordinate classification head transforms the image features into N one-dimensional vectors by one-dimensional convolution and full connection operation, representing the horizontal coordinate of each key point;

[0087] The visibility classification head transforms the image features into a one-dimensional vector with a length of N through two-dimensional convolution and full connection operation, representing the visibility of each point.

[0088] Further, the determination of the longitudinal coordinate of the key point through the longitudinal coordinate classification head, the determination of the transverse coordinate of the key point through the transverse coordinate classification head, and the determination of the visibility of the key point through the visibility classification head further comprise:

[0089] The longitudinal coordinate classification head first compresses the extracted key point features through two-dimensional convolution and ReLU activation function; flattens the transverse coordinate direction features of the compressed features, performs full connection operation and ReLU activation function, and converts the feature dimension to a times of the height of the coating edge image; performs scale-invariant one-dimensional convolution operation in the transverse coordinate direction, to obtain the output result of the longitudinal coordinate classification head of the key point.

[0090] The transverse coordinate classification head first compresses the extracted key point features through two-dimensional convolution and ReLU activation function; next, the longitudinal coordinate direction features of the compressed features are flattened, full connection operation and ReLU activation function are performed, and the feature dimension is converted to a times of the width of the original image; finally, scale-invariant one-dimensional convolution operation in the longitudinal coordinate direction is performed, to obtain the output result of the transverse coordinate classification head of the key point.

[0091] The visibility classification head first compresses the extracted key point features through two two-dimensional convolutions, then performs global pooling, flattening, and full connection operation to obtain the output result of the visibility classification head of the key point.

[0092] Through the above steps, the positioning accuracy of the key point can reach the sub-pixel level, not only the detailed position of the key point at the edge position of the coating edge image can be obtained, but also the visibility of the key point can be predicted.

[0093] Step S300, a fully connected network classification model is constructed and trained until the loss function converges, and the longitudinal coordinate, transverse coordinate and visibility of the key point are input into the fully connected network classification model to extract the key point coordinates of the virtual edge, fusion edge and real edge at the edge of the coating edge image.

[0094] Preferably, the fully connected network classification model needs to move the predicted key point to the image edge according to different key point types when performing inference.

[0095] Specifically, the predicted key point needs to be moved from the padding pixel area to the corresponding image edge according to the category of the point (whether on the image edge, on which edge of the image) when the model performs inference.

[0096] The foregoing step performs pixel filling processing on the image edge, improves the feature discrimination degree near the annotation point, and improves the key point detection accuracy of the full connection network classification model. However, since the features of the pixel filling area are similar to the features of the edge position to be obtained, as shown in Figure 4 . The feature extraction network is a full convolution network. The full convolution network has strong spatial generalization ability due to the spatial translation invariance of the convolution kernel, which occasionally causes a larger prediction peak value of the horizontal coordinate classification head at the pixel filling position, resulting in classification error of the horizontal coordinate, affecting the edge grabbing effect, as shown in Figure 5 . Therefore, a full connection network is added to the horizontal and vertical coordinate classification heads, as shown in Figure 2 , to appropriately reduce the spatial generalization of the entire model, eliminate the occasional larger prediction peak value of the horizontal coordinate classification head at the pixel filling position, and improve the overall prediction effect of the model. The output result of the horizontal coordinate classification head after adding the full connection network is shown in Figure 6 , which solves the problem of over-detection of the image edge caused by the pixel filling operation due to poor edge point detection effect, and appropriately reduces the generalization of the model.

[0097] Finally, the key point vertical coordinates, horizontal coordinates, and visibility are input into the full connection network classification model for key point classification, and the key point coordinates of the virtual edge, fusion edge, and real edge at the edge of the coating edge image are extracted. The full connection network classification model is trained until the loss function converges.

[0098] Loss function: The loss function is a function used to measure the difference between the predicted results of a deep learning model and the actual results. During the training process, the loss function is usually used to evaluate the performance of the model, and the parameters of the model are updated through the backpropagation algorithm to minimize the value of the loss function. Common loss functions include Mean-Square Error (MSE), Cross Entropy Loss, and Log Loss. MSE is a loss function for regression problems, which calculates the square of the average difference between the predicted value and the actual value. Cross Entropy Loss is a loss function for classification problems, which calculates the cross entropy between the predicted value and the actual value. Log Loss is a loss function for binary classification problems, which calculates the logarithmic difference between the predicted value and the actual value. During the training process, the smaller the value of the loss function, the smaller the difference between the predicted results of the model and the actual results, and the better the performance of the model.

[0099] Step S400, input the vertical coordinates, horizontal coordinates, and visibility of the key points of the coating edge image into the trained full connection network classification model for detection and obtain the key point coordinates of the virtual edge, fusion edge, and real edge at the edge of the coating edge image. The positions of the virtual edge, fusion edge, and real edge are determined according to the key point coordinates.

[0100] According to the key point coordinates of the virtual edge, the fusion edge and the solid edge at the edge of the coating edge image, the positions of the virtual edge, the fusion edge and the solid edge to be grabbed are determined by using the principle of determining a straight line by two points. The information of the coating film size and shape of the coating is measured according to the positions of the virtual edge, the fusion edge and the solid edge, which is used for evaluating the quality and consistency of the coating, and is helpful for improving the process flow, improving the product quality and controlling the production cost.

[0101] Embodiment two

[0102] The second aspect of the application provides an AI coating edge grabbing processing system based on a full connection network classification model, comprising:

[0103] A processing module is configured to acquire a coating edge image and extract corresponding key point features.

[0104] An analysis module is configured to analyze the key point features by using convolution and full connection operations, and determine the longitudinal coordinates, the horizontal coordinates and the visibility of the key points through a classification head.

[0105] A training module is configured to build a full connection network classification model and train the full connection network classification model until a loss function converges, input the longitudinal coordinates, the horizontal coordinates and the visibility of the key points into the full connection network classification model, and extract the key point coordinates of the virtual edge, the fusion edge and the solid edge at the edge of the coating edge image.

[0106] A detection module is configured to input the longitudinal coordinates, the horizontal coordinates and the visibility of the key points of the coating edge image into the trained full connection network classification model for detection, and acquire the key point coordinates of the virtual edge, the fusion edge and the solid edge at the edge of the coating edge image, and determine the positions of the virtual edge, the fusion edge and the solid edge according to the key point coordinates.

[0107] Embodiment three

[0108] The embodiment of the application further provides a computer device for AI coating edge grabbing processing based on a full connection network classification model, and the device comprises:

[0109] At least one processor;

[0110] At least one memory configured to store at least one program;

[0111] When the at least one program is executed by the at least one processor, the at least one processor implements the method for AI coating edge grabbing processing based on a full connection network classification model according to any one of the above embodiments.

[0112] The contents in the above method embodiments are applicable to the system embodiments, the system embodiments specifically implement the functions of the above method embodiments, and achieve the same beneficial effects as the above method embodiments.

[0113] Embodiment Four

[0114] The application further provides a computer readable storage medium, wherein an AI coating edge grabbing processing program based on a full connection network classification model is stored on the computer readable storage medium, and the AI coating edge grabbing processing program based on the full connection network classification model, when executed by a processor, implements steps of the AI coating edge grabbing processing method based on the full connection network classification model.

[0115] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The processor is a control center of the AI coating edge grabbing processing system based on the improved synthetic biology experiment observation data analysis processing system, and is connected with various parts of the AI coating edge grabbing processing system based on the full connection network classification model through various interfaces and lines.

[0116] The memory can be used to store the computer program and / or modules, and the processor realizes various functions of the AI coating edge grabbing processing system based on the full connection network classification model by running or executing the computer program and / or modules stored in the memory, and calling data stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc. The data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state memory devices.

[0117] Although the description of the application has been quite detailed and particularly described several embodiments, it is not intended to be limited to any of these details or embodiments or any special embodiment, but should be considered to effectively cover the intended scope of the application by reference to the appended claims, taking into account the prior art to provide a broad interpretation of the claims. In addition, the application is described above in embodiments that the inventor can foresee, the purpose of which is to provide a useful description, and non-essential modifications to the application that have not yet been foreseen can still represent equivalent modifications of the application.

[0118] Compared with the prior art, the beneficial effects of the application at least include:

[0119] The application provides an AI coating edge grabbing processing method and system based on a full connection network classification model, which combines traditional algorithms and advanced machine learning algorithms, improves the accuracy of edge grabbing recognition, and improves the robustness and stability of recognition. Compared with traditional algorithms, the subjective factors caused by individual differences and physical conditions of artificial are eliminated, the accuracy of recognition results is affected when the coating edge is uneven and the type is variable, and the disadvantages of difficult to identify virtual edge are eliminated. At the same time, the key point detection effect of the image edge of the artificial intelligence key point algorithm is not good, and the problem of easy over-detection of the image edge after filling is solved, effectively solving the problem of low accuracy of coating edge grabbing recognition.

[0120] The above-described embodiments only express several embodiments of the application, which are described in detail and in detail, but should not be understood as limiting the scope of the patent of the application. It should be noted that for ordinary skilled in the art, without departing from the concept of the application, a number of modifications and improvements can be made, which belong to the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. An AI coating edge-grabbing processing method based on a fully connected network classification model, characterized in that, Includes the following steps: Step S100: Obtain the coating edge image and extract the corresponding key point features; Step S200: Analyze the key point features using convolution and fully connected operations, and determine the ordinate, abscissa, and visibility of the key points using a classification head; Step S300: Construct a fully connected network classification model and train it until the loss function converges. Input the ordinate, abscissa, and visibility of the key points into the fully connected network classification model to extract the key point coordinates of the virtual edge, fused edge, and solid edge at the edge of the coating edge image. Step S400: Input the ordinate, abscissa, and visibility of the key points of the coating edge image into the fully connected network classification model that has been trained to detect and obtain the key point coordinates of the virtual edge, fused edge, and solid edge at the edge of the coating edge image, and determine the position of the virtual edge, fused edge, and solid edge based on the key point coordinates. The process of acquiring the coating edge image and extracting the corresponding key point features includes: Obtain the coating edge image, and gradually fill the image edge with pixels of a specific color outward. The color of each ring of pixels is the same, and the chromaticity difference and saturation difference between adjacent rings are set within a specific range. After filling the pixels, the keypoint labels located at the image edge in the image annotation data are moved to the outermost circle of the filled pixel area and then input into the subsequent steps for keypoint feature extraction.

2. The AI ​​coating edge-grabbing processing method based on a fully connected network classification model according to claim 1, characterized in that, The process of acquiring the coating edge image and extracting the corresponding key point features includes the following steps: The original image of the coating is acquired and preprocessed to convert the image into a uniform format. The preprocessing includes one or more of the following methods: scaling, cropping, grayscale enhancement, and contrast enhancement. The first edge position of the preprocessed coating original image is determined by the edge extraction algorithm, and the coating original image is cropped with the first edge position as the center to generate the coating edge image; The coating edge image is labeled with data, and a feature extraction network with a pre-trained model is used to extract key point features of the coating edge image.

3. The AI ​​coating edge-grabbing processing method based on a fully connected network classification model according to claim 1, characterized in that, The analysis of keypoint features using convolution and fully connected operations, and the determination of the keypoint's ordinate, abscissa, and visibility using a classification head, includes: The classification head includes a vertical axis classification head, a horizontal axis classification head, and a visibility classification head; The vertical coordinates of key points are determined by the vertical coordinate classification head, the horizontal coordinates of key points are determined by the horizontal coordinate classification head, and the visibility of key points is determined by the visibility classification head.

4. The AI ​​coating edge-grabbing processing method based on a fully connected network classification model according to claim 3, characterized in that, The process of determining the ordinate of a key point using a ordinate classification head, determining the abscissa of a key point using a lateral classification head, and determining the visibility of a key point using a visibility classification head includes: The ordinate classification head first compresses the extracted keypoint features using two-dimensional convolution and ReLU activation function; then flattens the compressed features in the horizontal direction, performs a fully connected operation and ReLU activation function, transforming its feature dimension to α times the height of the painted edge image; finally, it performs a scale-invariant one-dimensional convolution operation in the horizontal direction to obtain the output result of the ordinate classification head for the keypoints. The horizontal coordinate classification head first compresses the extracted keypoint features using two-dimensional convolution and ReLU activation function; then it flattens the compressed features in the vertical direction, performs a fully connected operation and ReLU activation function, transforming its feature dimension to α times the original image width; finally, it performs a scale-invariant one-dimensional convolution operation in the vertical direction to obtain the output result of the horizontal coordinate classification head for the keypoints. The visibility classification head first compresses the feature length and width by performing two 2D convolutions on the extracted key point features, and then performs global pooling, flattening and fully connected operations to obtain the visibility classification head output of the key points.

5. The AI ​​coating edge-grabbing processing method based on a fully connected network classification model according to claim 1, characterized in that, Include: When performing inference, the fully connected network classification model needs to move the predicted key points to the image edge according to different key point types. The key points extracted by the fully connected network classification model include the coordinates of two key points at the edge of the image for the virtual edge, the real edge, and the fused edge. The positions of the virtual edge, the fused edge, and the real edge are determined based on the principle that two points determine a straight line.

6. An AI coating edge-grabbing processing system based on a fully connected network classification model, comprising: The processing module is used to acquire the coating edge image and extract the corresponding key point features; The process of acquiring the coating edge image and extracting the corresponding key point features includes: Obtain the coating edge image, and gradually fill the image edge with pixels of a specific color outward. The color of each ring of pixels is the same, and the chromaticity difference and saturation difference between adjacent rings are set within a specific range. After filling the pixels, the key point labels located at the image edge in the image annotation data are moved to the outermost circle of the filled pixel area and then input into the subsequent steps for key point feature extraction. The analysis module uses convolution and fully connected operations to analyze the key point features, and uses a classification head to determine the ordinate, abscissa, and visibility of the key points. The training module constructs a fully connected network classification model and trains it until the loss function converges. The ordinates, abscissas, and visibility of the key points are input into the fully connected network classification model to extract the coordinates of key points at the edges of the painted edge image for virtual edges, fused edges, and solid edges. The detection module inputs the ordinates, abscissas, and visibility of key points in the coated edge image into the fully connected network classification model that has been trained to detect and obtain the coordinates of key points at the edges of the coated edge image for virtual edges, fused edges, and solid edges. The positions of virtual edges, fused edges, and solid edges are determined based on the key point coordinates.

7. A computer device, characterized in that, It includes a processor and a memory, the memory being used to store executable instructions for controlling the processor to perform the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 5.

Images