Defect AI visual sorting system of plastic product production line

Abstract

The invention relates to the field of plastic product production quality detection, and discloses a defect AI visual sorting system of a plastic product production line, comprising: a data acquisition unit for shooting plastic products on the production line by using an industrial camera array to generate original image data; performing dynamic background removal processing on the original image data to obtain main body image data; according to the invention, through the AI visual system, the industrial camera and the deep learning model, high-precision and automatic defect detection and sorting are realized, the sorting speed and accuracy of a production line are improved, manual intervention is reduced, and the production efficiency is improved. The system adopts advanced background removal, noise filtering and image enhancement technologies, ensures that only effective foreground information is processed, improves the robustness and accuracy of subsequent defect recognition, and particularly has outstanding performance in a complex environment.

Description

Technical Field

[0001] This invention relates to the field of quality inspection technology for plastic products, specifically to an AI visual sorting system for defects in plastic product production lines. Background Technology

[0002] Plastic products are widely used in daily life and industrial production, and their quality directly affects the performance and market competitiveness of downstream products. Common defects in the production process of plastic products include bubbles, scratches, black spots, breakage, and color differences. These defects may originate from mold problems, raw material impurities, unstable process parameters, or environmental factors.

[0003] Currently, the detection and sorting of defects in plastic products on production lines mainly rely on manual quality inspection, which presents the following problems: high labor costs, as production lines expand, manual inspection of each item is time-consuming and labor-intensive, and highly dependent on skilled workers; insufficient efficiency and consistency, as manual inspection is easily affected by fatigue and subjective judgment, leading to a high rate of missed defects and misjudgments; low level of automation, although some companies use simple optical inspection devices, they can only detect obvious defects and cannot accurately classify and grade complex defects; and disconnect between the sorting and inspection processes, as current inspection and sorting are often separate, making it difficult to provide timely feedback on inspection results to the sorting process, resulting in reduced production efficiency.

[0004] With the development of deep learning and computer vision, introducing artificial intelligence technology into the quality inspection and sorting of plastic products can realize closed-loop control of the entire process, including defect detection, defect classification, size measurement, severity assessment and automatic sorting, which can greatly improve production efficiency and product consistency. Summary of the Invention

[0005] The purpose of this invention is to provide an AI visual sorting system for defects in plastic product production lines, in order to solve the problems of high labor costs, insufficient efficiency and consistency, low level of intelligence and disconnection in sorting processes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an AI visual sorting system for defects in plastic product production lines, comprising: The data acquisition unit uses an industrial camera array to photograph plastic products on the production line, generating raw image data; it then performs dynamic background removal processing on the raw image data to obtain the main image data. The feature extraction unit performs multi-channel color space conversion on the main image data to obtain feature image data; based on the feature image data, it performs preliminary extraction of product surface texture features to generate preliminary texture feature data. The edge enhancement unit performs product region segmentation based on preliminary texture feature data to obtain region segmentation image data; it then performs edge enhancement processing on the region segmentation image data to generate edge-enhanced region image data. The defect type classification unit uses a convolutional neural network model to perform deep learning recognition of defect features on edge-enhanced region image data to obtain defect feature recognition result data; based on the defect feature recognition result data, preliminary defect type classification is performed to generate preliminary defect classification data. The sorting strategy unit measures the defect size from the preliminary defect classification data to generate defect size data; it combines the defect size data and the preliminary defect classification data to assess the severity of defects and generate defect severity assessment data; and it sets the sorting strategy based on the defect severity assessment data to generate sorting instruction data. The sorting unit transmits sorting instruction data to the intelligent sorting robotic arm control system on the production line. The sorting robotic arm grabs and sorts plastic products according to the sorting instructions, placing products of different grades into designated areas to complete the AI ​​visual sorting of defects in the plastic product production line.

[0007] Furthermore, the sorting strategy includes: Severely defective products are marked for Level 1 sorting; Products with general defects are marked for secondary sorting; Products without defects are marked as qualified products.

[0008] Furthermore, the data acquisition unit includes: An industrial camera array is deployed on a plastic product production line to capture images of the products on the production line at a set frequency, generating raw image data. The optical flow dynamic background modeling technique in OpenCV is used to analyze the original image data, construct a dynamic background model, and obtain dynamic background model data. The original image data is subjected to background subtraction operation using dynamic background model data to separate the foreground product image. Then, morphological processing is performed to remove noise interference to obtain the main image data.

[0009] Furthermore, the feature extraction unit includes: The main image data is converted from the RGB color space to the HSV and Lab color spaces to obtain multi-channel color feature image data. In each color space channel, the local binary mode algorithm is used to extract texture features from the multi-channel color feature image data to obtain preliminary texture feature data for multiple channels. The preliminary texture feature data from multiple channels are fused together, and a weighted average method is used to combine the features of each channel to generate the final preliminary texture feature data.

[0010] Furthermore, the edge strengthening unit includes: Based on preliminary texture feature data, an adaptive threshold segmentation algorithm is used to segment the plastic product image into regions. The adaptive threshold is calculated by the local mean method. The image is divided into different functional regions according to the differences in texture features to obtain region segmentation image data. The Sobel operator is used to perform edge detection on the region segmentation image data to obtain the edge information of each region. Then, the edges are enhanced by image dilation and erosion operations to improve the clarity of the boundaries of each region and generate edge-enhanced region image data.

[0011] Furthermore, the defect type classification unit includes: The edge enhancement region image data is input into a pre-trained convolutional neural network model. The convolutional neural network model extracts the defect features in the image through multi-layer convolution and pooling operations to obtain the defect feature recognition result data. Based on the defect feature identification results, the support vector machine classification algorithm is used to perform preliminary classification of defect types and generate preliminary defect classification data.

[0012] Furthermore, the defect type classification unit also includes: On the edge enhancement region image data, a pixel distance-based measurement method is used to measure the defect size in the preliminary defect classification data, obtain the pixel size of the defect in the image, and then combine it with the camera calibration parameters to convert it into actual size information, generating defect size data.

[0013] Furthermore, the defect type classification unit also includes: A defect severity assessment model is constructed, taking defect size data and preliminary defect classification data as input parameters. The model scores and assesses the severity of defects according to preset rules, generating defect severity assessment data.

[0014] Furthermore, the sorting strategy also includes: Different threshold ranges are set based on the defect severity assessment data. The assessment value corresponding to severely defective products is set within the first-level threshold range and marked as first-level sorting; the assessment value corresponding to moderately defective products is set within the second-level threshold range and marked as second-level sorting; the assessment value corresponding to non-defective products is set within the qualified threshold range and marked as qualified products. Based on different sorting marks, sorting instruction data is generated.

[0015] Furthermore, the sorting unit includes: The sorting instruction data is transmitted to the intelligent sorting robotic arm control system on the production line via wired or wireless communication. The sorting robot arm control system parses the sorting instruction data to obtain product location information and sorting level information; Based on the product location information, the robotic arm is controlled to move to the designated position. Then, based on the sorting grade information, the appropriate gripping method and sorting action are selected to accurately grip and place products of different grades into the designated area, thus completing the defect AI vision sorting of the plastic product production line.

[0016] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention uses an AI vision system, an industrial camera and a deep learning model to achieve high-precision and automated defect detection and sorting, thereby improving the sorting speed and accuracy of the production line and reducing manual intervention. (2) The system adopts advanced background removal, noise filtering and image enhancement technologies to ensure that only effective foreground information is processed and to improve the robustness and accuracy of subsequent defect identification, especially in complex environments; (3) By combining deep learning and support vector machines, the system can accurately identify various defect types and intelligently adjust the sorting strategy according to the severity of the defects, thereby efficiently classifying products and improving the automation level of the production line; (4) By reducing manual inspection and scrap rate, the system reduces production costs while improving product quality and production line flexibility, supporting changes and expansions in future production needs. Attached Figure Description

[0017] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a flowchart of the defect type classification unit of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-2 This invention provides a technical solution: an AI visual sorting system for defects in plastic product production lines, comprising: The data acquisition unit uses an industrial camera array to photograph plastic products on the production line, generating raw image data; it then performs dynamic background removal processing on the raw image data to obtain the main image data. The industrial camera array consists of an array structure composed of multiple industrial cameras arranged in a specific layout. The raw image data is a collection of unprocessed image information about plastic products captured by the industrial camera array. This data contains the original features of the product, such as its appearance, color, and shape, but may also contain irrelevant information such as the background. Dynamic background removal is an image processing technique designed to remove the background that changes over time from the raw image data. The main image data is the image data containing only the main body of the plastic product after dynamic background removal, providing a cleaner and more accurate data foundation for subsequent feature extraction and other operations. The feature extraction unit performs multi-channel color space conversion on the main image data to obtain feature image data; based on the feature image data, it performs preliminary extraction of product surface texture features to generate preliminary texture feature data. Multi-channel color space conversion is the process of converting the main image data from one color space to another color space with multiple channels; feature image data is the image data obtained after multi-channel color space conversion. At this point, each channel of the image contains information that is more conducive to analyzing product features, providing more targeted data for subsequent texture feature extraction; preliminary extraction of product surface texture features uses image processing algorithms to extract texture feature information of the product surface from the feature image data; preliminary texture feature data is a set of data about the texture characteristics of the product surface obtained through the preliminary extraction of product surface texture features. This data is represented in a certain form and contains basic information about the product surface texture; The edge enhancement unit performs product region segmentation based on preliminary texture feature data to obtain region segmentation image data; it then performs edge enhancement processing on the region segmentation image data to generate edge-enhanced region image data. Product region segmentation involves using image segmentation algorithms to separate the product from the background based on preliminary texture feature data, and further dividing the product into different regions. Region segmentation image data is the image data obtained after product region segmentation, where the product is clearly divided into different regions, each with relatively independent features, facilitating subsequent specific processing for different regions. Edge enhancement processing uses image enhancement algorithms to process the region segmentation image data, highlighting the edge information of the product regions. Edge-enhanced region image data is the image data obtained after edge enhancement processing, where the edges of the product regions are significantly enhanced, making the details and defect edges of the product surface more prominent, providing a clearer image basis for subsequent defect type classification. The defect type classification unit uses a convolutional neural network model to perform deep learning recognition of defect features on edge-enhanced region image data to obtain defect feature recognition result data; based on the defect feature recognition result data, preliminary defect type classification is performed to generate preliminary defect classification data. Among them, the convolutional neural network model is a deep learning model with characteristics such as local perception and weight sharing, which can automatically learn high-level features from image data. Defect feature deep learning recognition uses a trained convolutional neural network model to analyze and process image data of edge-enhanced regions. The model automatically extracts defect features from the image through operations such as convolution and pooling, and matches them with defect feature patterns learned during training, thereby identifying the defect features present in the image and obtaining defect feature recognition result data. The defect feature recognition result data is the data output by the convolutional neural network model after identifying defect features in the edge-enhanced region image data. It is usually represented in the form of probability or labels, indicating whether a specific type of defect exists in the image and related information about the defect, providing a basis for subsequent preliminary defect type classification. Preliminary defect type classification uses pre-set classification rules or algorithms based on the defect feature recognition result data to classify the identified defects into different types, such as scratches, bubbles, cracks, and color differences. The purpose of preliminary classification is to roughly categorize defects and provide basic information for the formulation of subsequent sorting strategies. Preliminary defect classification data is the data obtained after the preliminary defect type classification operation. It contains classification information of the type to which each detected defect belongs, usually represented in the form of labels or codes, which facilitates further processing and analysis of different types of defects. The sorting strategy unit measures the defect size from the preliminary defect classification data to generate defect size data; it combines the defect size data and the preliminary defect classification data to assess the severity of defects and generate defect severity assessment data; and it sets the sorting strategy based on the defect severity assessment data to generate sorting instruction data. Among them, defect size measurement uses image measurement algorithms to measure the size of defects in the preliminary defect classification data and obtain dimensional information such as the length, width, and area of ​​the defects. Defect size is a crucial indicator for assessing defect severity, and accurate size measurement is essential for developing reasonable sorting strategies. Defect size data is a quantitative set of data obtained after defect size measurement, representing the spatial size characteristics of defects and providing a key basis for subsequent defect severity assessment. Defect severity assessment combines defect size data and preliminary defect classification data, based on pre-set assessment standards and models, to comprehensively evaluate the severity of each defect. Defect severity assessment data, obtained after the defect severity assessment process, includes a quantitative assessment of the severity of each defect, typically expressed as a grade or score, providing a direct basis for setting subsequent sorting strategies. Sorting strategy setting involves developing corresponding sorting strategies based on the defect severity assessment data. Sorting instruction data, generated after the sorting strategy setting process, contains specific sorting operation information, such as the sorting target location, gripping method, and sorting sequence, guiding the intelligent sorting robotic arm control system on the production line to complete the sorting task. The sorting unit transmits sorting instruction data to the intelligent sorting robotic arm control system on the production line. The sorting robotic arm grabs and sorts plastic products according to the sorting instructions, placing products of different grades into designated areas to complete the defect AI visual sorting of plastic products production line. The intelligent sorting robotic arm control system is responsible for controlling the movement of the intelligent sorting robotic arm. It receives sorting instruction data from the sorting strategy unit and precisely controls the movement of each joint of the robotic arm according to the information in the instruction data, realizing the gripping, moving and placing of plastic products. The sorting robotic arm is a mechanical device with multiple degrees of freedom, which can move flexibly in three-dimensional space. The designated area is a pre-set area on the production line for storing plastic products of different grades. According to the sorting strategy, qualified products, slightly defective products and seriously defective products will be placed in different designated areas for subsequent processing, such as packaging, rework or scrapping.

[0020] It should be noted that during operation, the data acquisition unit utilizes an industrial camera array and performs dynamic background removal to accurately acquire images of the main product. The feature extraction unit uses multi-channel color space conversion and texture feature extraction to highlight key product features. The edge enhancement unit segments regions and enhances edges, making defects clearer and easier to identify. The defect type classification unit uses convolutional neural network deep learning to accurately identify and classify defects. The sorting strategy unit comprehensively assesses the severity of defects based on their size and type, and formulates a reasonable sorting strategy. The sorting unit operates precisely according to instructions, sorting products of different grades to designated areas, achieving efficient defect sorting and improving production efficiency and product quality. Sorting strategies include: Severely defective products are marked for Level 1 sorting; Products with general defects are marked for secondary sorting; Products without defects are marked as qualified products.

[0021] In one embodiment, the data acquisition unit includes: An industrial camera array is deployed on a plastic product production line to capture images of the products on the production line at a set frequency, generating raw image data. The optical flow dynamic background modeling technique in OpenCV is used to analyze the original image data, construct a dynamic background model, and obtain dynamic background model data. The original image data is subjected to background subtraction operation using dynamic background model data to separate the foreground product image. Then, morphological processing is performed to remove noise interference to obtain the main image data.

[0022] This design, by deploying an industrial camera array on the plastic product production line and shooting at a set frequency, can comprehensively and dynamically capture product images, ensuring that no key information is missed. The model is constructed using the optical flow dynamic background modeling technology in OpenCV, which can accurately distinguish between dynamic background and product foreground. The foreground product image is separated by background difference operation, and noise is removed by morphological processing, resulting in clear and accurate subject image data. This effectively copes with dynamic changes on the production line, eliminates background interference, provides high-quality images for subsequent processing, improves the accuracy and reliability of defect detection, reduces misjudgments caused by complex backgrounds, and enhances the performance of the entire vision sorting system.

[0023] In one embodiment, the feature extraction unit includes: The main image data is converted from the RGB color space to the HSV and Lab color spaces to obtain multi-channel color feature image data. In each color space channel, the local binary mode algorithm is used to extract texture features from the multi-channel color feature image data to obtain preliminary texture feature data for multiple channels. LBP is a commonly used algorithm in texture analysis. It generates a binary pattern by comparing the gray values ​​of neighboring pixels with those of the center pixel. The formula for calculating the LBP value is: in, The grayscale value of the surrounding pixels. The grayscale value of the center pixel. The number of neighboring pixels. For symbolic functions, it is defined as: ; The preliminary texture feature data from multiple channels are fused together, and a weighted average method is used to combine the features of each channel to generate the final preliminary texture feature data.

[0024] This design converts the main image data from RGB to HSV and Lab color spaces, enabling the extraction of richer feature information from the product across different color spaces. The use of a local binary mode algorithm to extract texture features in each channel effectively captures subtle texture variations on the product surface. Furthermore, weighted averaging and fusion of the multi-channel preliminary texture feature data integrates the advantages of each channel, generating more comprehensive and accurate preliminary texture feature data. This multi-dimensional extraction of product features enhances their expressiveness and discriminative power, facilitating more accurate identification of product defects and improving the comprehensiveness and effectiveness of feature extraction, thus laying a solid foundation for subsequent defect classification.

[0025] In one embodiment, the edge reinforcement unit includes: Based on preliminary texture feature data, an adaptive threshold segmentation algorithm is used to segment the plastic product image into regions. The adaptive threshold is calculated by the local mean method. The image is divided into different functional regions according to the differences in texture features, and the region segmentation image data is obtained. The Sobel operator is used to perform edge detection on the region segmentation image data to obtain the edge information of each region. Then, the edges are enhanced by image dilation and erosion operations to improve the clarity of the boundaries of each region and generate edge-enhanced region image data.

[0026] This design, based on preliminary texture feature data, employs an adaptive threshold segmentation algorithm for region segmentation. It can accurately divide different functional areas according to the differences in product texture features, adapting to the complex and varied surface conditions of products. By using the Sobel operator to detect edges and enhance them through dilation and erosion, the boundaries of each region can be clearly presented. The precise segmentation of product areas and the enhancement of edges make the product outline and defect boundaries clearer, facilitating the accurate identification of defect locations and shapes in the subsequent process. This improves the accuracy and stability of edge detection, enhances the system's ability to capture product edge details, and improves the accuracy of defect detection.

[0027] In one embodiment, the defect type classification unit includes: The edge enhancement region image data is input into a pre-trained convolutional neural network model. The convolutional neural network model extracts the defect features in the image through multi-layer convolution and pooling operations to obtain the defect feature recognition result data. Based on the defect feature identification results, the support vector machine classification algorithm is used to perform preliminary classification of defect types and generate preliminary defect classification data.

[0028] This design inputs edge-enhanced region image data into a pre-trained convolutional neural network model. Through multi-layer convolution and pooling operations, deep-level defect features of the image can be automatically extracted, resulting in accurate defect feature recognition data. The support vector machine classification algorithm is then used to initially classify the defect types, quickly and accurately dividing defects into different categories. Combining the powerful feature extraction capabilities of the convolutional neural network and the efficient classification performance of the support vector machine, accurate identification and classification of defect types are achieved, improving the accuracy and efficiency of classification. This provides a reliable basis for subsequent sorting strategy development, ensuring that different types of defects are handled correctly.

[0029] In one embodiment, the defect type classification unit further includes: On the edge enhancement region image data, a pixel distance-based measurement method is used to measure the defect size in the preliminary defect classification data, obtain the pixel size of the defect in the image, and then combine it with the camera calibration parameters to convert it into actual size information, generating defect size data.

[0030] This design employs a pixel-distance-based measurement method to measure defect dimensions in edge-enhanced region image data. By first acquiring the pixel dimensions and then converting them into actual size information using camera calibration parameters, the actual size of the defect can be accurately determined. Precise defect size measurement provides crucial quantitative data for assessing defect severity. Different defect sizes have varying impacts on product quality, and accurate size measurement helps to more scientifically determine the severity of defects, making sorting strategies more rational, ensuring product sorting meets quality requirements, and improving the overall quality level of products.

[0031] In one embodiment, the defect type classification unit further includes: A defect severity assessment model is constructed, taking defect size data and preliminary defect classification data as input parameters. The model scores and assesses the severity of defects according to preset rules, generating defect severity assessment data.

[0032] This design constructs a defect severity assessment model that uses defect size data and preliminary defect classification data as input parameters for scoring and evaluation. It comprehensively considers the combined impact of defect type and size on product quality, establishing a scientific and reasonable defect severity assessment system. This avoids the one-sidedness of assessment based on a single factor. Through model scoring, the severity of defects can be assessed more accurately and objectively, providing a more precise basis for sorting strategies. This ensures that products with defects of different severity are handled appropriately, improving the accuracy and effectiveness of product quality control and enhancing product market competitiveness.

[0033] In one embodiment, the sorting strategy further includes: Different threshold ranges are set based on the defect severity assessment data. The assessment value corresponding to severely defective products is set within the first-level threshold range and marked as first-level sorting; the assessment value corresponding to moderately defective products is set within the second-level threshold range and marked as second-level sorting; the assessment value corresponding to non-defective products is set within the qualified threshold range and marked as qualified products. Based on different sorting marks, sorting instruction data is generated.

[0034] This design, by setting different threshold ranges and marking sorting levels based on defect severity assessment data, clearly distinguishes between products with severe defects, general defects, and no defects. It establishes clear and detailed sorting standards, making the sorting process more targeted and operable. Different treatment methods are adopted for products of different levels, such as scrapping products with severe defects, reworking products with general defects, and directly packaging products without defects, thereby improving production efficiency and reducing resource waste. Clear sorting markings facilitate the generation of accurate sorting instruction data, guiding the sorting robot arm to accurately execute sorting tasks and ensuring that the sorting work is carried out efficiently and orderly.

[0035] In one embodiment, the sorting unit includes: The sorting instruction data is transmitted to the intelligent sorting robotic arm control system on the production line via wired or wireless communication. The sorting robot arm control system parses the sorting instruction data to obtain product location information and sorting level information; Based on the product location information, the robotic arm is controlled to move to the designated position. Then, based on the sorting grade information, the appropriate gripping method and sorting action are selected to accurately grip and place products of different grades into the designated area, thus completing the defect AI vision sorting of the plastic product production line.

[0036] This design, using wired or wireless communication to transmit sorting instruction data, enables rapid and stable transmission of instructions to the intelligent sorting robotic arm control system. The control system parses the instructions to obtain product location and sorting level information, then controls the robotic arm to accurately move, grasp, and place the products. This achieves rapid and accurate transmission and parsing of sorting instructions, ensuring the robotic arm can operate precisely according to the instructions. Wired communication is stable and reliable, while wireless communication is flexible and convenient, allowing for the selection of the appropriate method based on the actual conditions of the production line. The robotic arm accurately sorts products based on detailed information, improving sorting accuracy and efficiency, reducing manual intervention, lowering labor intensity, and achieving automation and intelligence in defect sorting of plastic product production lines.

[0037] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. An AI vision sorting system for defects in plastic product production lines, characterized in that, include: The data acquisition unit uses an industrial camera array to photograph plastic products on the production line and generate raw image data; The original image data is dynamically background removed to obtain the main image data. The data acquisition unit includes: arranging an industrial camera array on the plastic product production line to capture images of the products on the production line at a set frequency, generating original image data; using the optical flow dynamic background modeling technology in OpenCV to analyze the original image data, constructing a dynamic background model, and obtaining dynamic background model data; using the dynamic background model data to perform background subtraction on the original image data to separate the foreground product image, and then performing morphological processing to remove noise interference to obtain the main image data. The feature extraction unit performs multi-channel color space conversion on the main image data to obtain feature image data; based on the feature image data, it performs preliminary extraction of product surface texture features to generate preliminary texture feature data; the feature extraction unit includes: converting the main image data from RGB color space to HSV and Lab color spaces to obtain multi-channel color feature image data; using the local binary mode algorithm to extract texture features from the multi-channel color feature image data in each color space channel to obtain preliminary texture feature data for multiple channels; and fusing the preliminary texture feature data for multiple channels, using a weighted average method to combine the features of each channel to generate the final preliminary texture feature data; An edge enhancement unit performs product region segmentation based on preliminary texture feature data to obtain region segmentation image data; it then performs edge enhancement processing on the region segmentation image data to generate edge-enhanced region image data. The edge enhancement unit includes: segmenting the plastic product image into regions using an adaptive threshold segmentation algorithm based on preliminary texture feature data. The adaptive threshold is calculated using the local mean method. The image is divided into different functional regions based on texture feature differences to obtain region segmentation image data. The Sobel operator is used to perform edge detection on the region segmentation image data to obtain edge information for each region. Finally, image dilation and erosion operations are used to enhance the edges, improving the clarity of each region's boundaries and generating edge-enhanced region image data. The defect type classification unit uses a convolutional neural network model to perform deep learning recognition of defect features on edge-enhanced region image data to obtain defect feature recognition result data; based on the defect feature recognition result data, it performs preliminary defect type classification to generate preliminary defect classification data; the defect type classification unit also includes: constructing a defect severity assessment model, taking defect size data and preliminary defect classification data as input parameters, and the model scores and assesses the severity of defects according to preset rules to generate defect severity assessment data; The sorting strategy unit measures the defect size from the preliminary defect classification data to generate defect size data; it combines the defect size data and the preliminary defect classification data to assess the severity of the defects and generate defect severity assessment data; and it sets the sorting strategy based on the defect severity assessment data to generate sorting instruction data. The sorting unit transmits sorting instruction data to the intelligent sorting robotic arm control system on the production line. The sorting robotic arm grabs and sorts plastic products according to the sorting instructions, placing products of different grades into designated areas to complete the AI ​​visual sorting of defects in the plastic product production line.

2. The defect AI vision sorting system for plastic product production lines according to claim 1, characterized in that: The sorting strategy includes: Severely defective products are marked for Level 1 sorting; Products with general defects are marked for secondary sorting; Products without defects are marked as qualified products.

3. The defect AI vision sorting system for plastic product production lines according to claim 2, characterized in that: The defect type classification unit includes: The edge enhancement region image data is input into a pre-trained convolutional neural network model. The convolutional neural network model extracts the defect features in the image through multi-layer convolution and pooling operations to obtain the defect feature recognition result data. Based on the defect feature identification results, the support vector machine classification algorithm is used to perform preliminary classification of defect types and generate preliminary defect classification data.

4. The defect AI vision sorting system for plastic product production lines according to claim 3, characterized in that: The defect type classification unit also includes: On the edge enhancement region image data, a pixel distance-based measurement method is used to measure the defect size in the preliminary defect classification data, obtain the pixel size of the defect in the image, and then combine it with the camera calibration parameters to convert it into actual size information, generating defect size data.

5. The defect AI vision sorting system for plastic product production lines according to claim 4, characterized in that: The sorting strategy also includes: Different threshold ranges are set based on the defect severity assessment data. The assessment value corresponding to severely defective products is set within the first-level threshold range and marked as first-level sorting; the assessment value corresponding to moderately defective products is set within the second-level threshold range and marked as second-level sorting; the assessment value corresponding to non-defective products is set within the qualified threshold range and marked as qualified products. Based on different sorting marks, sorting instruction data is generated.

6. The defect AI vision sorting system for plastic product production lines according to claim 5, characterized in that: The sorting unit includes: The sorting instruction data is transmitted to the intelligent sorting robotic arm control system on the production line via wired or wireless communication. The sorting robot arm control system parses the sorting instruction data to obtain product location information and sorting level information; Based on the product location information, the robotic arm is controlled to move to the designated position. Then, based on the sorting grade information, the appropriate gripping method and sorting action are selected to accurately grip and place products of different grades into the designated area, thus completing the defect AI vision sorting of the plastic product production line.

Images