Intelligent inspection method and system based on AI image recognition

Abstract

The application provides an intelligent inspection method and system based on AI image recognition, and relates to the technical field of image recognition analysis. Wherein, the application firstly acquires visible light images, thermal imaging images and quality evaluation data of an inspection area; secondly, dynamically selects and combines a processing path from multiple preset processing paths and simultaneously performs feature improvement on the visible light images to generate a standardized feature map; then combines the standardized feature map with a temperature feature map to generate a combined feature map; finally, inputs the combined feature map into a pre-trained image classification model to output a determination result of food production process compliance. The technical scheme provided by the application forms a technical chain of sensing quality, optimizing features, precise fusion and stable decision-making, which improves the determination performance of intelligent inspection of process compliance in a complex real environment.

Description

Technical Field

[0001] This application relates to the field of image recognition and analysis technology, and in particular to an intelligent inspection method and system based on AI image recognition. Background Technology

[0002] With the increasing demands for standardization of production processes and food safety control in the catering and food processing industries, there is an urgent need for an intelligent inspection technology that can automatically and accurately monitor key parameters of food production processes to replace traditional manual inspections that are subjective, inefficient, and difficult to sustain.

[0003] Currently, the existing technical solution adopts an intelligent monitoring method based on multimodal image fusion. This method simultaneously acquires visible light images and thermal imaging images, registers and fuses them to form comprehensive data containing appearance texture and temperature information, and then directly inputs it into a pre-trained image classification model to determine whether the current process status meets the preset standards. This method achieves the joint utilization of visible appearance and temperature information to a certain extent.

[0004] However, this method typically uses a fixed image processing flow and feature fusion method, making it difficult to adapt to complex and ever-changing real industrial environments. Furthermore, when there are drastic changes in ambient light, slight camera defocus leading to image blurring, or temporary interference with the thermal imaging sensor, the fixed image preprocessing and feature extraction strategies struggle to generate stable and reliable feature representations, resulting in a decline in the quality of the data input to the classification model. Additionally, the method often neglects the direct impact of image quality on the fusion effect and the differentiated impact of different quality dimensions on the final decision, thus resulting in insufficient accuracy and robustness in judgments under non-ideal conditions, and a tendency for false positives or false negatives. Summary of the Invention

[0005] This application provides an intelligent inspection method and system based on AI image recognition to solve the problem that the existing technology cannot dynamically optimize the processing and fusion strategy according to the real-time image quality, resulting in low reliability of process conformity judgment in complex environments.

[0006] Firstly, this application provides an intelligent inspection method based on AI image recognition, including:

[0007] Acquire visible light and thermal images of the inspection area, as well as quality assessment data;

[0008] Based on the quality evaluation data, processing paths are dynamically selected and combined from multiple preset processing paths. The visible light image is then improved according to the combined processing paths to generate a standardized feature map.

[0009] The standardized feature map is combined with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map;

[0010] The combined feature map is input into a pre-trained image classification model, and the model outputs a judgment result on the conformity of the food production process based on the combined feature map.

[0011] Optionally, visible light and thermal images of the inspection area, as well as quality assessment data, are acquired, including:

[0012] Visible light images reflecting details of the scene in the inspection area and thermal images reflecting the temperature distribution in the inspection area are simultaneously acquired by imaging devices pre-deployed in the inspection area.

[0013] By analyzing the visible light image, global illumination parameters and regional sharpness parameters describing the visible light image are obtained;

[0014] By analyzing the thermal imaging images, temperature fluctuation characteristics that characterize the stability of temperature readings are extracted from the thermal imaging images;

[0015] The global illumination parameters, the regional sharpness parameters, and the temperature fluctuation characteristics are integrated to generate quality evaluation data.

[0016] Optionally, based on the quality evaluation data, processing paths are dynamically selected and combined from multiple preset processing paths, and the visible light image is feature-improved according to the combined processing paths to generate a standardized feature map, including:

[0017] The global illumination parameters in the quality evaluation data are compared with multiple preset illumination intensity threshold ranges. Based on the illumination intensity threshold range in which the global illumination parameters are located, a target image brightness adjustment path is selected from a preset image brightness adjustment path sequence.

[0018] The region sharpness parameter in the quality evaluation data is compared with a preset sharpness discrimination threshold. If the region sharpness parameter is lower than the sharpness discrimination threshold, a preset image sharpening enhancement path is activated.

[0019] The temperature fluctuation characteristics in the quality evaluation data are compared with a preset temperature stability threshold. Based on the changing trend of the temperature fluctuation characteristics relative to the temperature stability threshold, a target noise suppression path is selected from a preset noise suppression path sequence.

[0020] The target image brightness adjustment path, the preset image sharpening enhancement path (determined based on the judgment result), and the target noise suppression path are combined into an image processing path according to a preset execution order.

[0021] According to the processing method defined in each path of the image processing path, the visible light image is sequentially subjected to image brightness adjustment, image sharpening enhancement and noise suppression operations to obtain the processed visible light image;

[0022] A feature extraction operation is performed on the processed visible light image to generate a standardized feature map, which is used to describe the scene details of the inspection area.

[0023] Optionally, the standardized feature map is combined with a temperature feature map from the thermal imaging image at the same time to generate a combined feature map, including:

[0024] The thermal imaging image is segmented into regions. Based on the continuity and differences in the temperature distribution in the thermal imaging image, multiple temperature-related regions are divided, wherein each temperature-related region contains a set of pixels with similar temperature characteristics.

[0025] Calculate the characteristic temperature values ​​that represent the temperature level of the temperature-related region.

[0026] Establish a mapping relationship between the location of each temperature-related region in the thermal imaging image and the corresponding region in the standardized feature map;

[0027] Based on the mapping relationship, the characteristic temperature value corresponding to each temperature-related region is assigned to each feature point in the corresponding location region of the standardized feature map;

[0028] Using each feature point in the standardized feature map as a reference, the detailed feature information contained in the feature point is associated with the assigned feature temperature value to form a target feature point that carries both detailed features and temperature features.

[0029] All target feature points are aggregated to form a combined feature map.

[0030] Optionally, the combined feature map is input into a pre-trained image classification model, and a judgment result on the conformity of the food preparation process is output based on the combined feature map, including:

[0031] The combined feature map is input into a pre-trained image classification model;

[0032] In the image classification model, multiple process semantic levels are defined based on different process stages associated with food production processes;

[0033] Based on the detailed feature information and feature temperature value carried by each target feature point in the combined feature map, within each process semantic level, the target feature point is classified into a corresponding process semantic category, and the first confidence level of the target feature point belonging to the process semantic category is determined.

[0034] Based on the first confidence of all target feature points classified into different process semantic categories within the process semantic hierarchy, the second confidence of each process semantic category at the process semantic hierarchy is calculated.

[0035] The process semantic category with the highest second confidence level in each process semantic level is determined as the first judgment result of the process semantic level.

[0036] By combining the first determination results of all process semantic levels, a process compliance determination vector is generated.

[0037] Based on the process compliance determination vector, the final determination result for the compliance of the food production process is output.

[0038] Optionally, characteristic temperature values ​​representing the temperature level of the temperature-related region are calculated, including:

[0039] Obtain the set of pixel temperature values ​​for all pixels within the temperature-related region;

[0040] In the set of pixel temperature values, all pixel temperature values ​​are arranged in order from highest to lowest temperature to obtain a pixel temperature value sequence.

[0041] Remove a first preset proportion of pixel temperature values ​​from both ends of the pixel temperature value sequence to obtain a subset of pixel temperature values ​​in the middle region;

[0042] Calculate the arithmetic mean of all pixel temperature values ​​within the subset of pixel temperature values ​​in the intermediate region to obtain the initial feature temperature value;

[0043] The core temperature subset is formed by identifying all pixel temperature values ​​in the intermediate region whose difference from the initial feature temperature value is within a second preset range.

[0044] The feature temperature value is calculated based on the distribution pattern of all pixel temperature values ​​in the core temperature subset.

[0045] Optionally, based on the first confidence of all target feature points classified into different process semantic categories within the process semantic hierarchy, a second confidence of each process semantic category at the process semantic hierarchy is calculated, including:

[0046] The number of all target feature points classified into the same process semantic category within the stated process semantic level is counted.

[0047] The first confidence scores of all target feature points belonging to the same process semantic category are summed to obtain the cumulative confidence score of the process semantic category;

[0048] Calculate the ratio of the cumulative confidence score to the number to obtain the initial hierarchical confidence score of the process semantic category;

[0049] Determine the range of variation of the first confidence level for all target feature points contained in the process semantic category;

[0050] Based on the range of change, a target confidence correction strategy is selected from a preset set of confidence correction strategies;

[0051] Using a target confidence correction strategy, the initial level confidence is numerically adjusted to obtain a second confidence of the process semantic category at the process semantic level.

[0052] Secondly, this application provides an intelligent inspection system based on AI image recognition, comprising:

[0053] The acquisition module is used to acquire visible light images and thermal imaging images of the inspection area, as well as quality evaluation data.

[0054] An improvement module is used to dynamically select and combine processing paths from multiple preset processing paths based on the quality evaluation data, and to improve the features of the visible light image according to the combined processing paths to generate a standardized feature map.

[0055] A combination module is used to combine the standardized feature map with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map;

[0056] The output module is used to input the combined feature map into a pre-trained image classification model and output a judgment result on the conformity of the food production process based on the combined feature map.

[0057] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement an intelligent inspection method based on AI image recognition as described in the first aspect above.

[0058] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements an intelligent inspection method based on AI image recognition as described in the first aspect.

[0059] This application addresses the problem of unstable feature extraction in complex environments such as illumination changes and focus deviations by dynamically selecting and combining processing paths based on image quality evaluation data. It intelligently selects and combines the current optimal image enhancement strategy from a variety of preset processing paths based on real-time acquired multi-dimensional quality data such as illumination, sharpness, and temperature stability, thereby adaptively generating more robust standardized features and providing reliable input for subsequent analysis.

[0060] Furthermore, an adaptive correction mechanism based on confidence distribution is introduced in the classification decision stage to improve the overall reliability of process judgment. After obtaining the first confidence of each feature point, not only is the average confidence of each category calculated, but also the corresponding correction strategy is selected for adjustment based on the range of confidence variation within each category to obtain a more robust second confidence. This mechanism can effectively suppress the interference of local feature noise or outliers on the overall judgment, so that the final judgment result still maintains high credibility when there is uncertainty at the feature level.

[0061] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 A flowchart of an intelligent inspection method based on AI image recognition provided in this application is shown;

[0064] Figure 2 A schematic diagram of the structure of an intelligent inspection system based on AI image recognition provided in this application is shown;

[0065] Figure 3 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0066] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0067] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0068] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0069] Figure 1 This application provides a flowchart of an intelligent inspection method based on AI image recognition, such as... Figure 1 As shown, the method includes:

[0070] Step 101: Obtain visible light images and thermal images of the inspection area, as well as quality evaluation data.

[0071] Optionally, step 101 may specifically include:

[0072] Step 1011: Simultaneously acquire visible light images reflecting scene details of the inspection area and thermal imaging images reflecting the temperature distribution of the inspection area using an imaging device pre-deployed in the inspection area.

[0073] Step 1012: By analyzing the visible light image, obtain the global illumination parameters and regional sharpness parameters describing the visible light image.

[0074] Step 1013: Analyze the thermal imaging image to extract temperature fluctuation features that characterize the stability of temperature readings in the thermal imaging image.

[0075] Step 1014: Integrate the global illumination parameters, the regional sharpness parameters, and the temperature fluctuation characteristics to generate quality evaluation data.

[0076] In this step, visible light images refer to ordinary photographs taken by camera equipment installed in the inspection area, which record the color and brightness details of objects and are used to reflect the visual appearance information of the monitored scene.

[0077] Thermal imaging images are images captured by thermal imagers that record the intensity of infrared radiation on the surface of an object and convert it into temperature distribution, used to reflect the temperature field information of the monitored scene.

[0078] Quality assessment data refers to a set of indicators used to quantitatively evaluate whether the currently acquired image data is clear, stable, and suitable for subsequent analysis, and is used to guide the selection of subsequent image processing strategies.

[0079] Global illumination parameter refers to a value used to describe the average brightness of the entire visible light image. It is used to determine whether the ambient light is too bright or too dark, and is obtained by calculating the average brightness of all pixels in the image.

[0080] The regional sharpness parameter is a numerical value used to describe the degree of blur in a pre-defined key target area in an image, such as a stove or worktable. It is used to determine whether the image details are clear and usable, and is obtained by calculating the high-frequency component energy of the image in that area.

[0081] Temperature fluctuation characteristics refer to a numerical value used to describe the degree of drastic change in temperature readings at the same location in multiple consecutive thermal imaging images within a short period of time. It is used to determine whether thermal imaging data is subject to transient interference and is obtained by calculating the standard deviation of temperature values ​​in a specific area across multiple images.

[0082] In this step, firstly, visible light images containing scene details of the inspection area and thermal imaging images reflecting the temperature distribution of the inspection area are simultaneously acquired by imaging devices pre-deployed in the inspection area. Specifically, under the same command, visible light cameras and thermal imagers installed in fixed positions are simultaneously triggered to ensure that both capture the scene at the same time, thereby obtaining image pairs that are strictly corresponding in time.

[0083] Secondly, by analyzing the acquired visible light images, global illumination parameters and regional sharpness parameters are obtained. First, the color visible light image is converted into a grayscale image, and then the arithmetic mean of the brightness values ​​of all pixels in the grayscale image is calculated as the global illumination parameter. At the same time, according to the preset coordinates of key areas related to food processing, the corresponding sub-image areas are cropped from the original visible light image, and the Laplacian operator is applied to the sub-image areas for convolution operation. The standard deviation of the operation result is calculated, and this standard deviation value is used as the regional sharpness parameter of the inspection area.

[0084] Next, by analyzing the synchronously acquired thermal imaging images, the temperature fluctuation characteristics that characterize the stability of temperature readings are extracted, and the most recent few frames of thermal imaging images are continuously cached. A fixed monitoring area is selected, and for each pixel in the monitoring area, the standard deviation of its temperature value sequence in consecutive frames is calculated. Finally, the standard deviations of all pixels are averaged, and the average value is the current temperature fluctuation characteristic.

[0085] Finally, the calculated global illumination parameters, regional sharpness parameters, and temperature fluctuation characteristics are assembled into a multi-dimensional data vector according to a predefined format. This data vector is then used to generate quality evaluation data to guide subsequent processes.

[0086] For example, in workshop B of the frying production line of a food processing plant A, a visible light camera C and a thermal imager D are deployed and simultaneously capture images under command. First, the visible light camera C captures a color image containing the frying basket and the oil pan, while the thermal imager D simultaneously captures the temperature distribution map of the corresponding area. Second, after converting the color image captured by C into a grayscale image, the overall average brightness is calculated to be 120 (assuming an 8-bit grayscale range). At the same time, the central area of ​​the oil pan is selected according to a preset coordinate frame, and the clarity score of the central area is calculated to be 15.6 using the Laplacian operator. On the other hand, the temperature images captured by D in the last 5 seconds are read, and the fluctuation of the temperature reading in the oil pan area is calculated, resulting in an average temperature fluctuation characteristic value of 0.8 degrees Celsius. Finally, the values ​​120, 15.6, and 0.8 are integrated into a data vector 120, 15.6, 0.8, which serves as the quality evaluation data for the current moment.

[0087] Step 102: Based on the quality evaluation data, dynamically select and combine processing paths from multiple preset processing paths, and improve the features of the visible light image according to the combined processing paths to generate a standardized feature map.

[0088] Optionally, step 102 may specifically include:

[0089] Step 1021: Compare the global illumination parameters in the quality evaluation data with multiple preset illumination intensity threshold ranges, and select the target image brightness adjustment path from the preset image brightness adjustment path sequence according to the illumination intensity threshold range in which the global illumination parameters are located.

[0090] Step 1022: Compare the regional sharpness parameter in the quality evaluation data with a preset sharpness discrimination threshold. If the regional sharpness parameter is lower than the sharpness discrimination threshold, then activate the preset image sharpening enhancement path.

[0091] Step 1023: Compare the temperature fluctuation characteristics in the quality evaluation data with a preset temperature stability threshold, and select a target noise suppression path from a preset noise suppression path sequence based on the changing trend of the temperature fluctuation characteristics relative to the temperature stability threshold.

[0092] Step 1024: Combine the target image brightness adjustment path, the preset image sharpening enhancement path (determined based on the judgment result), and the target noise suppression path into an image processing path according to a preset execution order.

[0093] Step 1025: According to the processing method defined in each path of the image processing path, perform image brightness adjustment, image sharpening enhancement and noise suppression operations on the visible light image in sequence to obtain the processed visible light image.

[0094] Step 1026: Perform feature extraction on the processed visible light image to generate a standardized feature map, which is used to describe the scene details of the inspection area.

[0095] In this step, the preset processing path refers to a set of pre-designed operation instructions that correct or enhance a specific quality problem in the image, and is used to perform targeted processing on the image under different imaging conditions.

[0096] Standardized feature maps refer to image feature data that has undergone normalization and has a unified mathematical expression form. They are used to describe the key visual information of an image and serve as a stable input for subsequent analysis.

[0097] Global illumination parameter is a quantitative value that describes the overall brightness of an image. It is used to determine whether the ambient lighting is suitable and is obtained by calculating the average brightness of all pixels in the visible light image.

[0098] The preset multiple light intensity threshold intervals refer to a pre-defined set that divides the range of light intensity values ​​into several continuous segments, used to classify continuous light values ​​into different light levels.

[0099] A preset image brightness adjustment path sequence refers to a series of pre-arranged processing methods for adjusting image brightness. For example, path A brightens dark areas, path B suppresses overexposure, and path C keeps the image as is. These methods are used to select the appropriate brightness adjustment method based on different lighting levels.

[0100] The target image brightness adjustment path refers to the brightness adjustment method selected from the preset sequence based on the current lighting conditions, which is used to perform the most appropriate brightness correction on the current image.

[0101] The preset sharpness threshold is a numerical boundary used to determine whether an image is sharp enough, and is used to decide whether image sharpening processing needs to be enabled.

[0102] The preset image sharpening enhancement path refers to a set of predefined processing instructions used to enhance the sharpness of image edges and details, thereby improving the visual effect of blurry images.

[0103] The preset temperature stability threshold is a reference value used to determine whether the thermal imaging temperature readings are stable, and is used to assess the degree of interference with the thermal data.

[0104] The preset noise suppression path sequence refers to a series of pre-arranged different processing methods for suppressing random noise in an image, used to select the appropriate noise reduction method based on the degree of temperature instability.

[0105] The target noise suppression path refers to the noise reduction method actually selected from the preset sequence based on the current temperature fluctuation level, which is used to perform appropriate noise suppression on the image.

[0106] An image processing path refers to a complete operational flow that ultimately combines one or more specific processing steps to sequentially improve the original image.

[0107] The processed visible light image refers to the new visible light image obtained after all the operation steps in the image processing path, which is used to obtain an improved image quality.

[0108] Scene details in the inspection area refer to specific visual information in the image related to the monitored process, such as the shape of the equipment, the color of the ingredients, and the position of the tools. These are used to determine whether the process meets the specifications and are obtained by extracting key visual features from the processed image.

[0109] In this step, the global illumination parameter in the quality evaluation data is first read, and this global illumination parameter value is compared with multiple pre-stored illumination intensity threshold ranges. For example, three ranges, namely, too dark, normal, and too bright, and their corresponding numerical ranges, may be preset. By determining which numerical range the global illumination parameter falls into, the processing instruction bound to that illumination range is selected from another preset list, namely the image brightness adjustment path sequence, as the target image brightness adjustment path, and the brightness of the image is determined, such as whether to brighten the overall image or compress the highlights.

[0110] Next, the region sharpness parameter in the quality evaluation data is read and compared with another preset sharpness discrimination threshold. If the region sharpness parameter is determined to be lower than the discrimination threshold, it means that the key areas of the image are not sharp enough. Then, a preset image sharpening enhancement path will be activated. This image sharpening enhancement path contains specific sharpening filtering algorithms and parameters. Conversely, if the sharpness is sufficient, this path will not be activated.

[0111] Next, the temperature fluctuation characteristics in the quality evaluation data are read and compared with the preset temperature stability threshold. The trend of the change is analyzed to see if it is stable, slightly fluctuating or drastic. Based on the comparison result, the processing instruction that best matches the current temperature fluctuation level is selected from a preset noise suppression path sequence as the target noise suppression path. For example, a mild noise reduction algorithm is selected for slight fluctuations, and a stronger noise reduction algorithm is selected for drastic fluctuations.

[0112] Then, the three selection results are integrated to determine the target image brightness adjustment path to be used, whether to include the image sharpening enhancement path based on the sharpness comparison results, and the target noise suppression path to be used. These paths are combined into a complete image processing path to be executed in a preset fixed order, such as adjusting brightness first, then sharpening, and finally reducing noise.

[0113] Then the combined image processing path is executed. In the order defined in the path, the corresponding image processing algorithms are called on the original visible light image in turn. For example, the brightness adjustment algorithm is called first to correct the overall brightness of the image. Then, it is determined whether to call the sharpening filter algorithm to enhance the edges. Finally, the selected noise reduction algorithm is called to smooth the image noise. After all these operations, the processed visible light image with improved quality is obtained.

[0114] Finally, feature extraction is performed on the processed visible light image. Typically, a pre-trained deep convolutional neural network model is used. When the visible light image is input into the deep convolutional neural network, it will automatically extract high-level abstract features that can effectively represent the image content from its multi-layer structure. These high-level abstract features are organized into a standardized data block, namely a standardized feature map. This feature map describes the essential information of the scene details in the inspection area, removing redundancy and interference from the original pixels.

[0115] For example, following the specific implementation of the previous step, firstly, the quality evaluation data 120, 15.6, and 0.8 are obtained; secondly, the global illumination parameter 120 is extracted and compared with the preset range. It is found that 120 belongs to the normal illumination range, so the path of keeping the brightness unchanged is selected as the target image brightness adjustment path from the preset image brightness adjustment path sequence; next, the regional sharpness parameter 15.6 is extracted and compared with the sharpness threshold 10. Since 15.6 is higher than 10, it is determined that the image sharpness is sufficient, so it is decided not to enable the image sharpening enhancement path; then, the temperature fluctuation feature 0.8 is extracted and compared with the temperature stability threshold 1.0. Since 0.8 is lower than 1.0 and the fluctuation is small, the path of using a mild Gaussian filter is selected from the noise suppression path sequence. The target noise suppression path is then used; subsequently, the two paths of keeping the brightness unchanged and applying a light Gaussian filter are combined in a preset order to form the final image processing path; then, the original visible light image of the frying production line is processed according to this image processing path, first performing the operation of keeping the brightness unchanged, and then performing a light Gaussian filter to suppress possible weak noise, resulting in a cleaner processed visible light image; finally, the processed image is input into a pre-trained feature extraction network, which outputs a vector containing 512-dimensional features. This vector is the normalized feature map, which encodes key visual information about the oil pan, frying basket and their state.

[0116] This step dynamically selects and combines processing strategies based on real-time image quality data to adaptively optimize the image, improving its adaptability to complex field environments. Furthermore, by generating more stable standardized features, it provides a more robust input basis for subsequent process judgments.

[0117] Step 103: Combine the standardized feature map with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map.

[0118] Optionally, step 103 may specifically include:

[0119] Step 1031: Perform region segmentation on the thermal imaging image. Based on the continuity and differences in temperature distribution in the thermal imaging image, divide it into multiple temperature-related regions, wherein each temperature-related region contains a set of pixels with similar temperature characteristics.

[0120] Step 1032: Calculate the characteristic temperature values ​​representing the temperature level of the temperature-related region.

[0121] Optionally, step 1032 may specifically include the following steps: obtaining a set of pixel temperature values ​​for all pixels within the temperature-related region; arranging all pixel temperature values ​​in the set of pixel temperature values ​​according to the order of temperature from highest to lowest to obtain a pixel temperature value sequence; removing a first preset proportion of pixel temperature values ​​from both ends of the pixel temperature value sequence to obtain a subset of pixel temperature values ​​in the middle region; calculating the arithmetic mean of all pixel temperature values ​​in the subset of pixel temperature values ​​in the middle region to obtain an initial feature temperature value; determining all pixel temperature values ​​in the subset of pixel temperature values ​​in the middle region whose difference from the initial feature temperature value is within a second preset range to constitute a core temperature subset; and calculating a feature temperature value based on the distribution pattern of all pixel temperature values ​​in the core temperature subset.

[0122] Step 1033: Establish a mapping relationship between the position of each temperature-related region in the thermal imaging image and the corresponding position region in the standardized feature map.

[0123] Step 1034: Based on the mapping relationship, assign the characteristic temperature value corresponding to each temperature-related region to each feature point in the corresponding location region of the standardized feature map.

[0124] Step 1035: Using each feature point in the standardized feature map as a reference, associate the detailed feature information contained in the feature point with the assigned feature temperature value to form a target feature point that carries both detailed features and temperature features.

[0125] Step 1036: Aggregate all target feature points to form a combined feature map.

[0126] In this step, similar temperature features refer to the property that the temperature values ​​of multiple pixels in a thermal imaging image are close in value or have small variation. This feature is used to define a spatially continuous region with a high degree of consistency in temperature distribution. It is obtained by clustering and boundary division based on the temperature differences between pixels using an image segmentation algorithm.

[0127] Temperature feature maps refer to temperature information extracted from thermal imaging images and expressed in the form of structured data, used to characterize the temperature levels of different areas in the monitored scene.

[0128] Combined feature maps refer to a new data representation that integrates visual information from standardized feature maps with temperature information from temperature feature maps, providing multi-dimensional features that simultaneously include appearance and temperature for subsequent analysis.

[0129] Temperature-correlated regions refer to a continuous block in a thermal imaging image that is divided based on temperature similarity. It contains multiple pixels with similar temperature values ​​and is used to manage the temperature field in zones.

[0130] A pixel set refers to the collective term for all pixels that constitute a temperature-related region. These pixels are adjacent in image location and have similar temperature values, and are used to jointly define the temperature characteristics of the region. They are obtained by clustering through image segmentation algorithms.

[0131] Temperature level refers to a general description of the overall temperature of a region, used to quickly characterize the thermal state of the region, and is obtained by calculating the representative value of the pixel temperature value in the region.

[0132] The feature temperature value refers to a single value that can robustly represent the overall temperature level of a certain temperature-related region. It is used as a temperature label for that region in subsequent fusion and is obtained by statistical denoising and central tendency calculation of the pixel temperature values ​​in that region.

[0133] A pixel temperature value set refers to a list of the individual temperature values ​​of all pixels within a certain temperature-related region, used to reflect the original temperature distribution within that region.

[0134] A pixel temperature value sequence is an ordered list formed by arranging all the values ​​in the pixel temperature value set in descending or ascending order, and is used for order-based statistical processing.

[0135] The first preset ratio refers to a pre-set percentage value used to remove outliers that may exist at both ends of the sequence when calculating characteristic temperatures, such as removing the highest and lowest percentages of data.

[0136] A pixel temperature value subset refers to the middle set of data remaining after removing the data from both ends of the complete pixel temperature value sequence. It is used to focus on the main part of the temperature distribution and is obtained by cropping the sequence proportionally.

[0137] The initial feature temperature value refers to a preliminary representative value obtained by calculating a simple average of all values ​​within a subset of pixel temperature values. It is used as the basis for subsequent refined calculations and is obtained by calculating the arithmetic mean.

[0138] The second preset range refers to an interval centered on the initial characteristic temperature value, fluctuating up and down by a certain value, used to filter out core data that is highly consistent with the main body temperature.

[0139] The core temperature subset refers to the set of data in the pixel temperature value subset that falls within a second preset range. It is used to represent the most stable and consistent temperature information in that area and is obtained through numerical comparison and filtering.

[0140] Feature points refer to each basic data unit in a standardized feature map that represents local visual information of an image and is used to describe the detailed features of a point in a scene.

[0141] Detail feature information refers to multi-dimensional vector data stored in feature points, which is used to encode visual attributes such as texture, shape, and color of the corresponding image location, and is calculated through network forward propagation.

[0142] Target feature points refer to enhanced feature points that, after association processing, simultaneously contain detailed feature information and feature temperature values, and are used as the basic building blocks of combined feature maps.

[0143] In this step, the thermal imaging image is first divided into multiple temperature-related regions by applying a region-growing-based image segmentation algorithm. The image segmentation algorithm starts from a selected seed point and iteratively merges adjacent pixels whose temperature values ​​differ from the current region's average value by less than a set threshold, thereby obtaining a set of pixels with similar internal temperatures and spatial continuity.

[0144] Next, for each segmented temperature-related region, its characteristic temperature value is calculated, and the temperature values ​​of all pixels within the temperature-related region are obtained to form a set of pixel temperature values. A sorting algorithm is then used to arrange the set of pixel temperature values ​​from high to low, forming a sequence of pixel temperature values. At the same time, a predetermined proportion of data is removed from the beginning and end of the sequence to obtain a subset of pixel temperature values ​​in the middle region. The arithmetic mean of all values ​​in this subset is then calculated as the initial characteristic temperature value, and an allowable temperature difference threshold centered on this initial value is set. All values ​​falling within this range are selected from the subset of pixel temperature values ​​to form a core temperature subset. Finally, based on the distribution of the values ​​in the core subset, a weighted average is calculated to obtain the final characteristic temperature value.

[0145] Next, using pre-calibrated coordinate transformation parameters, a mapping relationship is established between the contour position of each temperature-related region in the thermal imaging image and the corresponding spatial range in the standardized feature map. Then, based on this mapping relationship, a temperature assignment operation is performed, and the feature temperature value calculated for each temperature-related region is assigned as an additional attribute to all feature points in the standardized feature map located within the corresponding mapping region.

[0146] Next, information association is performed, whereby the original multidimensional detailed feature information vector of each feature point is concatenated with the feature temperature value just assigned to it at the data structure layer to generate a target feature point that contains both types of information. Finally, through data aggregation, all the generated target feature points are reorganized according to their spatial relationship to form the final combined feature map.

[0147] For example, following the specific implementation of the previous step, firstly, a standardized feature map of the frying production line has been obtained; secondly, the synchronously captured thermal imaging image is processed, and through a region growing algorithm, the thermal imaging image is divided into two main temperature-related regions: one is a high-temperature region A centered on the oil pan, with a temperature range of 180-200 degrees Celsius; the other is the surrounding low-temperature background region B, with a temperature of approximately 30 degrees Celsius; then, for the high-temperature region A, the temperature values ​​of all its pixels are obtained, sorted, and the highest and lowest 5% of data are removed, the average value of the intermediate subset is calculated to be 190 degrees Celsius, and then filtered... Data within the range of 185 to 195 degrees Celsius were selected as the core subset, and their characteristic temperature value was calculated to be 191 degrees Celsius. Then, through coordinate mapping, the position of the high-temperature region A in the thermal imaging image was mapped to the feature points describing the visual features of the oil pan in the standardized feature map, and the value of 191 degrees Celsius was assigned to these points. Subsequently, each feature point describing the oil pan bound its original visual feature vector to the temperature value of 191, forming a target feature point. Finally, all target feature points, together with the feature points in the background area, constituted the final combined feature map that simultaneously contained information about the appearance of the oil pan and the oil temperature of 191 degrees Celsius.

[0148] This step achieves effective fusion of visual features and temperature information through robust temperature region segmentation and feature extraction, combined with precise spatial mapping. It transforms temperature data into stable labels associated with visual semantic regions, generating more comprehensive and strongly correlated multimodal combined features, thereby improving the system's ability to analyze process steps that depend on a combination of morphology and temperature.

[0149] Step 104: Input the combined feature map into the pre-trained image classification model, and output the judgment result of the conformity of the food production process based on the combined feature map.

[0150] Optionally, step 104 may specifically include:

[0151] Step 1041: Input the combined feature map into the pre-trained image classification model.

[0152] Step 1042: In the image classification model, multiple process semantic levels are defined based on different process stages associated with food production processes.

[0153] Step 1043: Based on the detailed feature information and feature temperature value carried by each target feature point in the combined feature map, within each process semantic level, the target feature point is classified into a corresponding process semantic category, and the first confidence level of the target feature point belonging to the process semantic category is determined.

[0154] Step 1044: Calculate the second confidence level of each process semantic category in the process semantic level based on the first confidence level of all target feature points classified into different process semantic categories within the process semantic level.

[0155] Optionally, step 1044 may specifically include the following steps: counting the number of all target feature points classified into the same process semantic category within the process semantic level; summing the first confidence scores of all target feature points belonging to the same process semantic category to obtain the cumulative confidence score of the process semantic category; calculating the ratio of the cumulative confidence score to the number to obtain the initial level confidence score of the process semantic category; determining the range of variation of the first confidence scores of all target feature points included in the process semantic category; selecting a target confidence score correction strategy from a preset set of confidence score correction strategies based on the range of variation; and using the target confidence score correction strategy to numerically adjust the initial level confidence score to obtain the second confidence score of the process semantic category at the process semantic level.

[0156] Step 1045: The process semantic category with the highest second confidence level in each process semantic level is determined as the first determination result of the process semantic level.

[0157] Step 1046: Combine the first judgment results of all process semantic levels to generate a process compliance judgment vector.

[0158] Step 1047: Based on the process compliance determination vector, output the final determination result for the compliance of the food production process.

[0159] In this step, the pre-trained image classification model refers to a deep neural network model that has already completed the learning process using a large amount of labeled food processing image data. It is used to analyze and understand the combined feature maps and output process-related judgments.

[0160] Process semantic hierarchy refers to multiple predefined analytical levels based on different stages or dimensions of interest in the food production process, such as raw material preparation, processing operations, and finished product status. These levels are used to evaluate the process in stages and from multiple perspectives and are defined through business knowledge.

[0161] Process semantic categories refer to the specific process states or result labels that are subdivided under each process semantic level. For example, under the processing operation level, they can be divided into categories such as normal temperature, excessively high temperature, and standard operation. These categories are used to describe the specific situations that may occur at this level and are defined through business rules.

[0162] The first confidence score refers to the probability score predicted by the image classification model for each target feature point in the combined feature map to belong to a certain process semantic category. It is used to measure the degree to which a single local feature point supports a certain process state and is calculated through the forward propagation of the network model.

[0163] The second confidence score refers to the overall confidence score of a category calculated by combining the first confidence scores of all target feature points classified into the same process semantic category within a certain process semantic level. It is used to evaluate the reliability of the category from the perspective of feature point groups.

[0164] The cumulative confidence score refers to the sum of the first confidence scores of all target feature points belonging to a certain process semantic category, which is used to reflect the cumulative support of that category.

[0165] The initial level confidence score refers to the average value obtained by dividing the sum of the cumulative confidence scores of a process semantic category by the number of target feature points belonging to that category. It is used to initially reflect the average confidence level of the category and is obtained through division.

[0166] The range of the first confidence level refers to the difference between the maximum and minimum values ​​of the first confidence level of all target feature points belonging to the same process semantic category. This difference is used to measure the dispersion of the confidence level within the group of points in that category and is calculated by statistically analyzing the maximum and minimum values.

[0167] The set of preset confidence correction strategies refers to a variety of pre-designed calculation rules or functions used to adjust the confidence value. For example, when the dispersion is large, a punitive downgrade is performed, while when the dispersion is small, it is maintained or fine-tuned. These strategies are used to optimize the initial confidence value based on the concentration of the confidence distribution.

[0168] The first judgment result refers to the process semantic category that is most likely to occur after comprehensive calculation and comparison within each independent process semantic level, and is used to represent the stage judgment conclusion of that level.

[0169] The process compliance determination vector is a vector composed of the first determination results of all process semantic levels arranged in order, used to comprehensively and structurally describe the compliance status of the entire process in various aspects.

[0170] In this step, the combined feature map is first input into a pre-trained deep convolutional neural network model through forward propagation. The deep convolutional neural network model performs nonlinear transformation and information abstraction on the input features through its multiple convolutional and fully connected layers. Secondly, during the inference process of the deep convolutional neural network model, multiple parallel fully connected classification layers are activated according to the predefined food process knowledge structure. Each classification layer corresponds to a process semantic level, such as the temperature control level or the finished product form level.

[0171] Next, within each process semantic level, each target feature point is processed by a Softmax classifier. The Softmax classifier calculates the probability distribution of the target feature point belonging to each process semantic category under that level based on the detailed feature information and feature temperature value of the target feature point, and selects the highest probability value from the probability distribution as the first confidence of the target feature point, and determines the corresponding category as its classification result.

[0172] Then, confidence aggregation and correction operations are performed. For each process semantic category, a counting algorithm is used to count the number of target feature points belonging to it, and an accumulation algorithm is used to sum the first confidence of these target feature points to obtain the cumulative confidence sum. Then, the initial level confidence is calculated by division. At the same time, an algorithm for finding the maximum and minimum values ​​is used to determine the range of variation of the first confidence of this type of target feature point group. Based on this range of variation, a target confidence correction strategy is selected from the preset rule set using a matching algorithm. For example, when the dispersion is large, an exponential decay function is applied for adjustment. Finally, the target confidence correction strategy is applied to recalculate the initial level confidence and output the second confidence.

[0173] Then, within each process semantic level, the maximum value of the second confidence score among all categories is found through a comparison and sorting algorithm, and the corresponding process semantic category is determined as the first judgment result of that process semantic level. Subsequently, through vector concatenation, the first judgment results of each process semantic level are combined in a preset order to generate a process compliance judgment vector. Finally, based on this process compliance judgment vector, the final judgment result is output through the decision logic module. For example, when all process semantic level judgments are compliant, the overall qualified conclusion is output.

[0174] For example, following the specific implementation of the previous step, firstly, the combined feature map containing information about the appearance of the oil pan and the temperature of 191 degrees Celsius is input into a pre-trained image classification model; secondly, the image classification model predefines two process semantic levels: oil temperature state and frying color; then, the image classification model first processes each target feature point. For example, for feature points describing the oil pan area, the first confidence level for determining whether it belongs to the "too high temperature" category under the oil temperature state level is 0.05, and the first confidence level for determining whether it belongs to the "normal temperature" category is 0.90.

[0175] Then, within the oil temperature level, assuming 95% of the feature points are classified into the normal temperature category, their average first confidence level is 0.88, but the confidence level varies considerably. Subsequently, based on this large variation range, a moderately downward adjustment strategy is selected, adjusting the second confidence level of the normal temperature category to 0.82, which is still the highest. Therefore, the first determination result of the oil temperature level is normal temperature. Similarly, the first determination result of the frying color level is golden color. Finally, these two results are combined into a vector: normal temperature, golden color, and the final output is the determination result that the current frying process meets the standard.

[0176] This step, based on feature point-level classification using deep networks, further corrects the overall judgment by aggregating local information and evaluating its internal consistency, effectively reducing local noise interference. The resulting structured, multi-dimensional judgment results improve the accuracy and reliability of intelligent inspection decisions.

[0177] Figure 2 This application provides a structural schematic diagram of an intelligent inspection system based on AI image recognition, as shown below. Figure 2 As shown, the system includes:

[0178] The acquisition module 21 is used to acquire visible light images and thermal imaging images of the inspection area, as well as quality evaluation data.

[0179] The improvement module 22 is used to dynamically select and combine processing paths from multiple preset processing paths based on the quality evaluation data, and to improve the features of the visible light image according to the combined processing paths to generate a standardized feature map.

[0180] The combination module 23 is used to combine the standardized feature map with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map;

[0181] The output module 24 is used to input the combined feature map into a pre-trained image classification model and output a judgment result on the conformity of the food production process based on the combined feature map.

[0182] Figure 2 The aforementioned AI-based image recognition-based intelligent inspection system can perform... Figure 1 The implementation principle and technical effects of the AI ​​image recognition-based intelligent inspection method described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the AI ​​image recognition-based intelligent inspection system in the above embodiments perform operations have been described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0183] In one possible design, Figure 2The AI ​​image recognition-based intelligent inspection system shown in the embodiment can be implemented as a computing device, such as... Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0184] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 32.

[0185] The processing component 32 is used for the above Figure 1 The embodiment describes an intelligent inspection method based on AI image recognition.

[0186] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above-described method. Alternatively, the processing component may be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described method.

[0187] Storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0188] Of course, computing devices may also include other components, such as input / output interfaces, display components, communication components, etc.

[0189] Input / output interfaces provide interfaces between processing components and peripheral interface modules, which can be output devices, input devices, etc.

[0190] The communication components are configured to facilitate wired or wireless communication between computing devices and other devices.

[0191] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform. In this case, the computing device can refer to a cloud server, and the aforementioned processing components, storage components, etc., can be basic server resources rented or purchased from the cloud computing platform.

[0192] This application also provides a computer storage medium storing a computer program, which, when executed by a computer, can perform the above-described functions. Figure 1The illustrated embodiment presents an intelligent inspection method based on AI image recognition.

[0193] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0194] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0195] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An intelligent inspection method based on AI image recognition, characterized in that, include: Acquire visible light and thermal images of the inspection area, as well as quality assessment data; Based on the quality evaluation data, processing paths are dynamically selected and combined from multiple preset processing paths. The visible light image is then improved according to the combined processing paths to generate a standardized feature map. The standardized feature map is combined with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map; The combined feature map is input into a pre-trained image classification model, and the model outputs a judgment result on the conformity of the food preparation process based on the combined feature map. The acquisition of visible light and thermal images of the inspection area, as well as quality evaluation data, includes: Visible light images reflecting details of the scene in the inspection area and thermal images reflecting the temperature distribution in the inspection area are simultaneously acquired by imaging devices pre-deployed in the inspection area. By analyzing the visible light image, global illumination parameters and regional sharpness parameters describing the visible light image are obtained; By analyzing the thermal imaging images, temperature fluctuation characteristics that characterize the stability of temperature readings are extracted from the thermal imaging images; The global illumination parameters, the regional sharpness parameters, and the temperature fluctuation characteristics are integrated to generate quality evaluation data; The process of dynamically selecting and combining processing paths from multiple preset processing paths based on the quality evaluation data, and performing feature improvement on the visible light image according to the combined processing paths to generate a standardized feature map, includes: The global illumination parameters in the quality evaluation data are compared with multiple preset illumination intensity threshold ranges. Based on the illumination intensity threshold range in which the global illumination parameters are located, a target image brightness adjustment path is selected from a preset image brightness adjustment path sequence. The region sharpness parameter in the quality evaluation data is compared with a preset sharpness discrimination threshold. If the region sharpness parameter is lower than the sharpness discrimination threshold, a preset image sharpening enhancement path is activated. The temperature fluctuation characteristics in the quality evaluation data are compared with a preset temperature stability threshold. Based on the changing trend of the temperature fluctuation characteristics relative to the temperature stability threshold, a target noise suppression path is selected from a preset noise suppression path sequence. The target image brightness adjustment path, the preset image sharpening enhancement path (determined based on the judgment result), and the target noise suppression path are combined into an image processing path according to a preset execution order. According to the processing method defined in each path of the image processing path, the visible light image is sequentially subjected to image brightness adjustment, image sharpening enhancement and noise suppression operations to obtain the processed visible light image; A feature extraction operation is performed on the processed visible light image to generate a standardized feature map, which is used to describe the scene details of the inspection area.

2. The method according to claim 1, characterized in that, The standardized feature map is combined with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map, including: The thermal imaging image is segmented into regions based on the continuity and differences in temperature distribution within the thermal imaging image. Multiple temperature-related regions are divided, each containing a set of pixels with similar temperature characteristics. These similar temperature characteristics refer to the property that the temperature values ​​of multiple pixels in the thermal imaging image are numerically close or have small variations. This property is used to define a spatially continuous region with a high degree of consistency in temperature distribution. The region is obtained by clustering and boundary delineation based on the temperature differences between pixels using an image segmentation algorithm. Calculate the characteristic temperature values ​​that represent the temperature level of the temperature-related region. Establish a mapping relationship between the location of each temperature-related region in the thermal imaging image and the corresponding region in the standardized feature map; Based on the mapping relationship, the characteristic temperature value corresponding to each temperature-related region is assigned to each feature point in the corresponding location region of the standardized feature map; Using each feature point in the standardized feature map as a reference, the detailed feature information contained in the feature point is associated with the assigned feature temperature value to form a target feature point that carries both detailed features and temperature features. All target feature points are aggregated to form a combined feature map.

3. The method according to claim 1, characterized in that, The combined feature map is input into a pre-trained image classification model, and a judgment result on the conformity of the food preparation process is output based on the combined feature map, including: The combined feature map is input into a pre-trained image classification model; In the image classification model, multiple process semantic levels are defined based on different process stages associated with food production processes; Based on the detailed feature information and feature temperature value carried by each target feature point in the combined feature map, within each process semantic level, the target feature point is classified into a corresponding process semantic category, and the first confidence level of the target feature point belonging to the process semantic category is determined. Based on the first confidence of all target feature points classified into different process semantic categories within the process semantic hierarchy, the second confidence of each process semantic category at the process semantic hierarchy is calculated. The process semantic category with the highest second confidence level in each process semantic level is determined as the first judgment result of the process semantic level. By combining the first determination results of all process semantic levels, a process compliance determination vector is generated. Based on the process compliance determination vector, the final determination result for the compliance of the food production process is output.

4. The method according to claim 2, characterized in that, Calculate characteristic temperature values ​​representing the temperature level of the temperature-correlated region, including: Obtain the set of pixel temperature values ​​for all pixels within the temperature-related region; In the set of pixel temperature values, all pixel temperature values ​​are arranged in order from highest to lowest temperature to obtain a pixel temperature value sequence. Remove a first preset proportion of pixel temperature values ​​from both ends of the pixel temperature value sequence to obtain a subset of pixel temperature values ​​in the middle region; Calculate the arithmetic mean of all pixel temperature values ​​within the subset of pixel temperature values ​​in the intermediate region to obtain the initial feature temperature value; The core temperature subset is formed by identifying all pixel temperature values ​​in the intermediate region whose difference from the initial feature temperature value is within a second preset range. The feature temperature value is calculated based on the distribution pattern of all pixel temperature values ​​in the core temperature subset.

5. The method according to claim 3, characterized in that, Based on the first confidence score of all target feature points classified into different process semantic categories within the process semantic hierarchy, a second confidence score for each process semantic category at the process semantic hierarchy is calculated, including: The number of all target feature points classified into the same process semantic category within the stated process semantic level is counted. The first confidence scores of all target feature points belonging to the same process semantic category are summed to obtain the cumulative confidence score of the process semantic category; Calculate the ratio of the cumulative confidence score to the number to obtain the initial hierarchical confidence score of the process semantic category; Determine the range of variation of the first confidence level for all target feature points contained in the process semantic category; Based on the range of change, a target confidence correction strategy is selected from a preset set of confidence correction strategies; Using a target confidence correction strategy, the initial level confidence is numerically adjusted to obtain a second confidence of the process semantic category at the process semantic level.

6. An intelligent inspection system based on AI image recognition, applied to the intelligent inspection method based on AI image recognition according to any one of claims 1-5, characterized in that, include: The acquisition module is used to acquire visible light images and thermal imaging images of the inspection area, as well as quality evaluation data. An improvement module is used to dynamically select and combine processing paths from multiple preset processing paths based on the quality evaluation data, and to improve the features of the visible light image according to the combined processing paths to generate a standardized feature map. A combination module is used to combine the standardized feature map with the temperature feature map from the thermal imaging image at the same time to generate a combined feature map; The output module is used to input the combined feature map into a pre-trained image classification model and output a judgment result on the conformity of the food production process based on the combined feature map.

7. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement an intelligent inspection method based on AI image recognition as described in any one of claims 1 to 5.

8. A computer storage medium, characterized in that, The system contains a computer program that, when executed by a computer, implements an intelligent inspection method based on AI image recognition as described in any one of claims 1 to 5.

Images