Al-based plastic toy visual inspection system and method

Abstract

The invention relates to the technical field of AI, and more specifically discloses an Al-based plastic toy visual inspection system and method, which can realize automatic quality inspection and classification identification work by acquiring a plastic toy image, extracting a feature vector, constructing a classification feature vector and utilizing a classifier to judge whether the plastic toy has defects. A traditional manual quality inspection method is replaced, and efficiency and accuracy are improved.

Description

Technical Field

[0001] The present application relates to the field of AI technology, and more specifically, to an AI-based plastic toy visual inspection system and method. Background Art

[0002] Behind the high-quality plastic toy products, there are increasingly complex processing, assembly and testing production processes, which puts higher and higher requirements on plastic toy AI visual inspection systems and automated machines. Plastic toy AI visual inspection is not only for quality inspection of plastic toys, it can also participate in most quality inspection work in industrial manufacturing. The application of AI visual inspection in industrial manufacturing is becoming more and more extensive, and plastic toys are just one of them. No matter what kind of industrial product, it must have quality assurance, and companies need to go through a series of defect detection quality before they can bring their products to market.

[0003] In factories, AI visual inspection will replace manual quality inspection and classification and identification work. Plastic toy AI visual inspection equipment will inevitably replace traditional manual quality inspection, which is not only inefficient but also has a low accuracy rate, not to mention that people in a fatigued emotional state may bring uncontrollable factors. Traditional manual quality inspection can only increase manpower to improve efficiency, and the training of a skilled quality inspector is not without cost. From recruitment, to employment, to getting started, and finally to proficiency, a lot of costs are spent in it. With the emergence of machine vision and artificial intelligence, more and more industrial assembly lines are now using online visual inspection products for appearance defects and classification identification.

[0004] Therefore, a plastic toy visual inspection system and method based on Al is desired. Summary of the invention

[0005] In order to solve the above technical problems, this application is proposed. The embodiment of this application provides an AI-based plastic toy visual inspection system and method, which uses the AI ​​visual inspection system to perform surface defect detection and classification identification on plastic toys, realizes automated quality inspection, and improves efficiency and accuracy.

[0006] Accordingly, according to one aspect of the present application, there is provided an Al-based plastic toy visual inspection system, comprising:

[0007] A plastic toy detection and acquisition module, used to obtain a plastic toy image of a plastic toy to be detected;

[0008] A plastic toy detection processing module, used to extract a plastic toy enhancement feature vector and a plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected;

[0009] A plastic toy detection fusion module, used for performing view consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector;

[0010] The plastic toy detection and analysis module is used to pass the plastic toy image classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be detected has defects.

[0011] According to another aspect of the present application, there is also provided a plastic toy visual inspection method based on Al, which comprises:

[0012] Acquire a plastic toy image of the plastic toy to be detected;

[0013] Extracting a plastic toy enhanced feature vector and a plastic toy image block feature vector from the plastic toy image to be detected;

[0014] Performing perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector;

[0015] The classification feature vector of the plastic toy image is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be inspected has defects.

[0016] Compared with the prior art, the present application provides an AI-based plastic toy visual inspection system and method, which acquires plastic toy images, extracts feature vectors and constructs classification feature vectors, and uses classifiers to determine whether the plastic toys have defects, thereby realizing automated quality inspection and classification identification, replacing traditional manual quality inspection methods, and improving efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0018] Figure 1 Schematic diagram of a block diagram of an Al-based plastic toy visual inspection system according to an embodiment of the present application.

[0019] Figure 2 Schematic diagram of a block diagram of a plastic toy detection processing module in an Al-based plastic toy visual detection system according to an embodiment of the present application.

[0020] Figure 3 Schematic diagram of a block diagram of a texture feature extraction unit in an Al-based plastic toy visual inspection system according to an embodiment of the present application.

[0021] Figure 4 It is a block diagram of a block feature extraction unit in an Al-based plastic toy visual inspection system according to an embodiment of the present application.

[0022] Figure 5 Flow chart of the Al-based plastic toy visual inspection method according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0024] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0025] In addition, in order to better illustrate the present application, numerous specific details are given in the following specific embodiments. It should be understood by those skilled in the art that the present application can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present application.

[0026] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0027] Figure 1 FIG. 1 is a block diagram of a plastic toy visual inspection system based on Al according to an embodiment of the present application. Figure 1As shown, according to the embodiment of the present application, the plastic toy visual inspection system 100 based on Al includes: a plastic toy detection acquisition module 110, which is used to obtain a plastic toy image of a plastic toy to be inspected; a plastic toy detection processing module 120, which is used to extract a plastic toy enhanced feature vector and a plastic toy image block feature vector from the plastic toy image of the plastic toy to be inspected; a plastic toy detection fusion module 130, which is used to perform a perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector; and a plastic toy detection analysis module 140, which is used to pass the plastic toy image classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be inspected has defects.

[0028] In the embodiment of the present application, the plastic toy detection acquisition module 110 is used to obtain the plastic toy image of the plastic toy to be detected. It should be understood that the purpose of obtaining the plastic toy image of the plastic toy to be detected is to provide the visual input data required by the system. Through these images, the system can perform surface defect detection and classification recognition, thereby judging the quality of the plastic toy. The monitoring image contains the visual information of the plastic toy, including features such as shape, color, and texture, which are essential for identifying and analyzing whether the plastic toy has defects. By extracting, processing, and analyzing the monitoring image, the system can identify the feature vector and image block feature matrix of the plastic toy, and then construct the plastic toy image classification feature vector. These feature vectors and feature matrices provide the data basis required for the system to perform quality inspection, which helps the system to accurately judge whether the plastic toy has defects. Therefore, the purpose of obtaining the plastic toy image of the plastic toy to be detected is to provide the input data required for the system to perform automated quality inspection. By processing and analyzing the image, the system can effectively judge the quality of the plastic toy and improve the efficiency and accuracy of the production line.

[0029] In the embodiment of the present application, the plastic toy detection processing module 120 is used to extract the plastic toy enhanced feature vector and the plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected. It should be understood that the purpose of extracting the plastic toy enhanced feature vector and the plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected is to extract feature information meaningful for quality detection so that the system can accurately determine whether the plastic toy has defects. The process of these feature extractions is to convert complex image data into a form that can be understood and processed by the computer, thereby realizing automated quality inspection. By extracting the enhanced feature vector of the plastic toy, the key features in the plastic toy image, such as color, shape, texture and other information, can be captured. These feature vectors can help the system better distinguish different types of plastic toys and identify potential defects or problems. By dividing the plastic toy image into small blocks and extracting the feature matrix, the local features of the image can be analyzed more carefully. Doing so helps the system to perform a more detailed analysis of the local area of ​​the plastic toy during the detection process, thereby improving the accuracy and sensitivity of the detection. By extracting these feature vectors and feature matrices, the system can establish an abstract representation of the plastic toy image, so that the system can more effectively perform defect detection and classification recognition. These feature extraction processes are a very important step in the quality inspection system. They provide the system with the necessary data basis, helping the system to accurately judge the quality of plastic toys and improve the efficiency and accuracy of automated quality inspection.

[0030] Specifically, in one embodiment of the present application, Figure 2 FIG. 1 is a block diagram of a plastic toy detection processing module in an Al-based plastic toy visual detection system according to an embodiment of the present application. Figure 2 As shown, in the above-mentioned Al-based plastic toy visual inspection system 100, the plastic toy detection processing module 120 includes: a texture feature extraction unit 121, used to extract texture features from the plastic toy image of the plastic toy to be inspected to obtain the plastic toy enhanced feature vector; a block feature extraction unit 122, used to extract the plastic toy image block feature vector from the plastic toy image of the plastic toy to be inspected.

[0031] Accordingly, in a specific example of the present application, the texture feature extraction unit 121 is used to extract texture features from the plastic toy image of the plastic toy to be detected to obtain the plastic toy enhanced feature vector. It should be understood that texture features are one of the important features for describing image details and local structures. By extracting texture features, the feature representation of the plastic toy image can be enriched, making the enhanced feature vector more representative and discriminative. The appearance texture of a plastic toy is usually one of its unique features, and different types of plastic toys may have different texture features. Therefore, extracting texture features helps to distinguish different types of plastic toys, thereby improving the system's recognition accuracy of plastic toys. Texture features usually have a certain stability to lighting changes and some transformations, which means that even under different lighting conditions or when there is a certain degree of deformation, texture features can still provide effective information. This stability helps the system maintain good performance in various environments. Therefore, extracting texture features to obtain a plastic toy enhanced feature vector can improve the robustness, accuracy and generalization ability of the system, thereby better supporting the quality detection and recognition tasks of plastic toys.

[0032] further, Figure 3 FIG. 1 is a block diagram of a texture feature extraction unit in an Al-based plastic toy visual inspection system according to an embodiment of the present application. Figure 3 As shown, in the plastic toy detection processing module 120 of the above-mentioned Al-based plastic toy visual inspection system 100, the texture feature extraction unit 121 includes: a grayscale co-occurrence matrix subunit 1211, which is used to extract a plastic toy texture feature map from the plastic toy image of the plastic toy to be detected based on the grayscale co-occurrence matrix; a cascade subunit 1212, which is used to cascade the plastic toy image with the plastic toy texture feature map to obtain a plastic toy multi-channel image; a residual dual attention subunit 1213, which is used to pass the plastic toy multi-channel image through a residual dual attention mechanism model to obtain a plastic toy enhanced multi-channel feature map; a plastic toy feature dimensionality reduction subunit 1214, which is used to perform global mean pooling on each feature matrix of the plastic toy enhanced multi-channel feature map along the channel dimension to obtain the plastic toy enhanced feature vector.

[0033] Specifically, the grayscale co-occurrence matrix subunit 1211 is used to extract a plastic toy texture feature map from the plastic toy image of the plastic toy to be detected based on the grayscale co-occurrence matrix. It should be understood that the grayscale co-occurrence matrix is ​​an effective method for describing image texture features. By analyzing the grayscale value relationship between pixels, rich texture information can be extracted, which is very useful for distinguishing different texture patterns and texture details. Plastic toys usually have a variety of different textures, such as granular, striped, etc. The grayscale co-occurrence matrix can capture the grayscale change pattern between local pixels in the image, thereby effectively characterizing the texture features and helping to distinguish different types of plastic toys. The grayscale co-occurrence matrix method has certain stability and robustness to illumination changes and some image noise. This means that even under imperfect illumination conditions or when there is some interference, the grayscale co-occurrence matrix can still provide effective texture features. The grayscale co-occurrence matrix is ​​a computationally simple and efficient method suitable for image processing tasks. Through the analysis of the grayscale co-occurrence matrix, effective texture features can be extracted in a shorter time, which helps the real-time performance and efficiency of the system. Therefore, extracting texture features from plastic toy images based on gray-level co-occurrence matrix is ​​an effective method, which can improve the system's ability to capture the texture features of plastic toys, thereby supporting the quality inspection and recognition tasks of plastic toys.

[0034] Specifically, the cascade subunit 1212 is used to cascade the plastic toy image with the plastic toy texture feature map to obtain a plastic toy multi-channel image. It should be understood that the cascade operation can combine the pixel information of the original image with the extracted texture feature information, thereby obtaining a more comprehensive and rich feature representation. This helps the system to better understand the image content. By cascading the texture feature map, the texture features of the image can be enhanced, making the feature representation more prominent and accurate. This helps to improve the system's recognition and classification capabilities for plastic toy images. The obtained plastic toy multi-channel image contains the color information of the original image and the enhanced information of the texture features. This multi-channel feature is more comprehensive and comprehensive, and can better describe the appearance characteristics of the plastic toy. By combining the original image and the texture features, the system can obtain a more informative and expressive feature representation, thereby improving the performance and accuracy of the plastic toy visual inspection system.

[0035] Specifically, the residual dual attention subunit 1213 is used to pass the plastic toy multi-channel image through the residual dual attention mechanism model to obtain a plastic toy enhanced multi-channel feature map. It should be understood that the residual connection can help the network better learn the residual information, thereby enhancing the representation ability of the feature. Combined with the attention mechanism, the network can pay more attention to important features, thereby further enhancing the feature representation ability of the plastic toy multi-channel image. The residual connection helps to alleviate the gradient disappearance problem, making it easier for the network to learn long-distance dependencies. The attention mechanism can help the network focus on important features at different positions, which helps to improve the network's perception range and learning ability. Through the residual connection and attention mechanism, the network's ability to extract and characterize plastic toy image features can be enhanced, which helps to improve the accuracy of quality detection. Such a mechanism can better capture the key features in the image, thereby improving the performance of the system. The residual connection and attention mechanism can improve the robustness of the model, making it more adaptable to changes in input data. This helps the model to stably extract effective features under different conditions. Therefore, processing the multi-channel images of plastic toys through the residual dual attention mechanism model can effectively enhance the feature representation ability, improve the performance and accuracy of the system, and at the same time improve the robustness and generalization ability of the model.

[0036] Correspondingly, the residual dual attention sub-unit includes: a spatial attention secondary sub-unit, which is used to pass the plastic toy multi-channel image through the spatial attention module of the residual dual attention mechanism model to obtain a plastic toy spatial attention map; a channel attention secondary sub-unit, which is used to pass the plastic toy multi-channel image through the channel attention module of the residual dual attention mechanism model to obtain a plastic toy channel attention map; a fusion weighted secondary sub-unit, which is used to calculate the plastic toy spatial attention map and the plastic toy channel attention map according to the position points to obtain a plastic toy weighted feature map; a fusion feature secondary sub-unit, which is used to fuse the plastic toy multi-channel image and the plastic toy weighted feature map to obtain the plastic toy enhanced feature map.

[0037] Specifically, the spatial attention secondary subunit is used to pass the plastic toy multi-channel image through the spatial attention module of the residual dual attention mechanism model to obtain a plastic toy spatial attention map. It should be understood that in image processing, spatial information is very important, and the relationship between pixel values ​​and pixels at different positions is crucial to understanding the image content. Through the spatial attention module, the model can learn the correlation between different positions, so as to better utilize spatial information. The spatial attention mechanism helps the network focus on important features at different positions when processing images, thereby improving the network's perception range. This enables the model to better understand the content at different positions in the image and helps to improve the representation ability of features. Through the spatial attention module, the model can pay more attention to the importance of different areas in the image, thereby enhancing the representation ability of features. This helps to improve the model's understanding and expression ability of image content. The spatial attention mechanism helps to reduce the impact of redundant information in the image and makes the model more focused on important spatial positions. This can improve the efficiency and accuracy of the model. Therefore, processing the plastic toy multi-channel image through the spatial attention module to obtain a plastic toy spatial attention map is to enhance the model's perception of spatial information and improve the representation ability of features, thereby improving the performance and accuracy of the system when processing plastic toy images.

[0038] Correspondingly, the spatial attention secondary subunit includes: a pooling tertiary subunit, which is used to perform average pooling and maximum pooling on the plastic toy multi-channel image along the channel dimension to obtain a plastic toy average feature matrix and a plastic toy maximum feature matrix; a cascading tertiary subunit, which is used to cascade and channel-adjust the plastic toy average feature matrix and the plastic toy maximum feature matrix to obtain a plastic toy channel feature matrix; a convolutional coding tertiary subunit, which is used to use the convolutional layer of the spatial attention module to convolutionally encode the plastic toy channel feature matrix to obtain a plastic toy convolutional feature matrix; a Softmax function tertiary subunit, which is used to pass the plastic toy convolutional feature matrix through the Softmax function to obtain a plastic toy spatial attention score matrix; and a positional point multiplication tertiary subunit, which is used to positionally point multiply the plastic toy spatial attention score matrix with each feature matrix of the plastic toy multi-channel image along the channel dimension to obtain the plastic toy spatial attention map.

[0039] Specifically, the channel attention secondary subunit is used to pass the plastic toy multi-channel image through the channel attention module of the residual dual attention mechanism model to obtain a plastic toy channel attention map. It should be understood that in deep learning, there may be a certain correlation between the features of different channels. Through the channel attention module, the model can learn the correlation between different channels, so as to better utilize the channel information. The channel attention mechanism helps to reduce the redundant information between channels in the image, so that the model pays more attention to the channels that are important to the current task. This helps to improve the efficiency and accuracy of the model. Through the channel attention module, the model can pay more attention to the importance of different channels in the image, thereby enhancing the representation ability of the features. This helps to improve the model's understanding and expression ability of the image content. By introducing the channel attention mechanism, the model can better adapt to different types of plastic toy image data, improve the generalization ability of the model, so that it performs more stably and reliably in different situations. Therefore, processing the plastic toy multi-channel image through the channel attention module to obtain the plastic toy channel attention map is to enhance the model's perception of channel features and improve the representation ability of features, thereby improving the performance and accuracy of the system when processing plastic toy images.

[0040] Correspondingly, the channel attention secondary subunit includes: a global pooling tertiary subunit, which is used to perform global mean pooling on each feature matrix along the channel dimension of the plastic toy multi-channel image to obtain a plastic toy channel feature vector; a normalization processing tertiary subunit, which is used to pass the plastic toy monitoring channel feature vector through a Softmax function to obtain a normalized plastic toy channel feature vector; a weighting tertiary subunit, which is used to weight the feature matrix along the channel dimension of the plastic toy multi-channel image using the eigenvalues ​​of each position in the normalized plastic toy channel feature vector as weights to obtain a plastic toy channel attention map.

[0041] Specifically, the fusion weighted secondary subunit is used to calculate the position points between the plastic toy spatial attention map and the plastic toy channel attention map to obtain the plastic toy weighted feature map. It should be understood that after calculating the plastic toy spatial attention map and the plastic toy channel attention map, the weighted feature map of the plastic toy can be obtained by element-wise multiplication of the two attention maps. The purpose of this step is to combine spatial information and channel information to better capture the relationship between different positions and different channels in the image, so as to obtain a more representative and rich feature representation. The spatial attention map captures the importance of different positions in the image, while the channel attention map focuses on the importance of different channels. By combining these two attention mechanisms, spatial and channel information can be comprehensively utilized to obtain a more comprehensive and rich feature representation. By multiplying the two attention maps by position, the spatial and channel information can be cross-influenced with each other, so that each position and channel in the feature map can be appropriately weighted, thereby improving the expressiveness and discrimination of the feature. Through the weighted feature map, the model can focus more on the features that are important to the current task, thereby improving the performance and accuracy of the model. Therefore, by point-wise multiplying the plastic toy spatial attention map and the plastic toy channel attention map by position, a weighted feature map of the plastic toy can be obtained, which can better capture the important information in the image and improve the performance of the model.

[0042] Specifically, the fusion feature secondary subunit is used to fuse the plastic toy multi-channel image and the plastic toy weighted feature map to obtain the plastic toy enhanced feature map. It should be understood that the plastic toy multi-channel image contains rich global information, while the plastic toy weighted feature map emphasizes important local information. By fusing the two, global and local information can be integrated to obtain a more comprehensive feature representation. Fusion of multi-channel images and weighted feature maps can enhance the expressiveness of features, making the features more distinguishable and representative. This helps to improve the model's understanding and classification performance of plastic toy images. The weighted feature map emphasizes important features through the attention mechanism, and after fusion, these key features can be combined with the information of the original image, so that the key features are more prominent and more influential. The obtained plastic toy enhanced feature map can be used as a more representative and informative input, which helps to improve the performance and accuracy of the model in tasks such as plastic toy recognition or classification. Therefore, the fusion of plastic toy multi-channel images and plastic toy weighted feature maps to obtain plastic toy enhanced feature maps is to improve the expressiveness of features, integrate global and local information, strengthen key features, and ultimately improve the performance and effect of the model in processing plastic toy image tasks.

[0043] Specifically, the plastic toy feature dimension reduction subunit 1214 is used to perform global mean pooling on each feature matrix along the channel dimension of the plastic toy enhanced multi-channel feature map to obtain the plastic toy enhanced feature vector. It should be understood that through global mean pooling, the channel information in the multi-channel feature map can be reduced to a feature vector, reducing the data dimension, thereby reducing the computational complexity and memory consumption. Global mean pooling can retain the important information of each channel and integrate it into a global feature vector. This helps to capture the global features of the entire image, not just the local features. Through global mean pooling, the influence of the spatial position information in the feature map on the feature representation can be reduced, making the features more robust and unaffected by position changes. Global mean pooling can make the network invariant to input images of different sizes because it integrates the information of all channels into a feature vector of fixed length. Therefore, by performing global mean pooling on each feature matrix of the plastic toy enhanced multi-channel feature map along the channel dimension, a plastic toy enhanced feature vector can be obtained, which helps to extract global features and reduce dimensions, providing more convenient and effective feature representation for subsequent tasks.

[0044] Accordingly, in a specific example of the present application, the block feature extraction unit 122 is used to extract the plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected. It should be understood that the plastic toy image block feature vector can help to divide the image into small local areas so as to better capture and represent the features of these local areas. This helps to improve the understanding of details and local structures in the plastic toy image. By extracting the image block feature matrix, the amount of data to be processed can be reduced, thereby reducing the computational complexity. This helps to improve the efficiency and speed of the algorithm. Extracting the image block feature matrix can enhance the local information in the image, so that the model is more focused on the feature representation of each image block, thereby better capturing subtle feature changes and structural information. By extracting image block feature matrices of different sizes and shapes, it is possible to adapt to target objects or image structures of different scales, and improve the model's detection ability for objects of different scales. Therefore, extracting the plastic toy image block feature vector from the plastic toy image to be detected helps to better capture and represent the local information in the image, reduce computational complexity, enhance local information, and adapt to target objects or image structures of different scales.

[0045] further, Figure 4 FIG. 1 is a block diagram of a block feature extraction unit in a plastic toy visual inspection system based on Al according to an embodiment of the present application. Figure 4As shown, in the plastic toy detection processing module 120 of the above-mentioned Al-based plastic toy visual inspection system 100, the block feature extraction unit 122 includes: an image block processing subunit 1221, which is used to divide the plastic toy image into image blocks to obtain a plastic toy image block sequence; a linear encoding subunit 1222, which is used to encode the plastic toy image block sequence through a linear embedding layer to obtain a plurality of plastic toy image block embedding vectors; a two-dimensional arrangement subunit 1223, which is used to arrange the plurality of plastic toy image block embedding vectors in two dimensions to obtain a plastic toy image block input matrix; a convolutional encoding subunit 1224, which is used to pass the plastic toy image block input matrix through a first convolutional neural network model with mutually transposed convolution kernels in adjacent layers to obtain a plastic toy image block feature matrix; a pooling subunit 1225, which is used to expand the plastic toy image block feature matrix into the plastic toy image block feature vector according to a row vector or a column vector.

[0046] Accordingly, the image block processing subunit 1221 is used to divide the plastic toy image into image blocks to obtain a sequence of plastic toy image blocks. It should be understood that by dividing the image into small image blocks, the local features in the image can be better captured. This helps to improve the understanding of details and local structures in the plastic toy image, thereby improving the accuracy of detection. Localized processing can make the system more robust to position changes in the image. Even if the position of the plastic toy in the image changes slightly, the system can still accurately identify and detect them. Processing small image blocks is more efficient than processing the entire image. By processing the image in blocks, the computational complexity can be reduced and the efficiency of the system can be improved. Dividing the image into blocks can make the system adapt to plastic toys of different sizes and shapes. Each image block can be processed independently, thereby better adapting to different plastic toy samples.

[0047] Accordingly, the linear encoding subunit 1222 is used to encode the plastic toy image block sequence through a linear embedding layer to obtain a plurality of plastic toy image block embedding vectors. It should be understood that through linear embedding layer encoding, the high-dimensional pixel data of each image block can be converted into a low-dimensional embedding vector, thereby achieving dimensionality reduction and abstraction of image information. This helps to reduce the complexity of data, improve computational efficiency, and better capture the key features of image blocks. Embedding vectors are usually designed as more representative feature representations that can better express the semantic information and features of image blocks. Such representations help improve the performance of subsequent processing steps, such as tasks such as classification and detection. Converting image blocks of different sizes into embedded vectors of fixed dimensions helps to achieve a unified feature representation, so that image blocks of different sizes can be processed in the same processing flow. Through this encoding method, the system can more effectively utilize the information of plastic toy image blocks and provide better input feature representations for subsequent processing steps.

[0048] Specifically, the linear encoding subunit is used to: pass the plastic toy image block sequence through an embedding layer to linearly embed code each plastic toy image block in the plastic toy image block sequence to obtain the multiple plastic toy image block embedding vectors, wherein the embedding layer uses a learnable embedding matrix to embed code each plastic toy image block.

[0049] Correspondingly, the two-dimensional arrangement subunit 1223 is used to arrange the multiple plastic toy image block embedding vectors in two dimensions to obtain a plastic toy image block input matrix. It should be understood that arranging the embedding vectors in the form of a matrix can better maintain the spatial information between the image blocks. In the plastic toy image, different blocks may represent different parts or features, and the relative position information can be better retained by arranging the matrix. Neural networks usually require input data to have a fixed shape. Arranging the embedding vectors in the form of a matrix can make it easier to dock with the neural network structure, especially for models such as convolutional neural networks that require two-dimensional input. In deep learning, data is usually input in a batch processing manner. Arranging the embedding vectors in a matrix form can more effectively perform batch processing operations and improve computational efficiency. By arranging the embedding vectors into a matrix, the implementation of subsequent processing steps can be simplified. For example, in a convolutional neural network, the input is usually two-dimensional, so arranging the embedding vectors into a matrix can be directly used as the input of the network to avoid additional conversion operations. By arranging multiple embedding vectors in a matrix form, models such as neural networks can be better used to process image data, improving the performance and efficiency of the model.

[0050] Correspondingly, the convolutional encoding subunit 1224 is used to input the plastic toy image block into the matrix through the adjacent layers using the first convolutional neural network model with mutually transposed convolution kernels to obtain the plastic toy image block feature vector. It should be understood that the convolutional neural network (CNN) is a powerful feature extractor that can effectively capture local features in an image through convolution operations. In this case, using mutually transposed convolution kernels can help the network learn richer and more abstract features, which helps to better represent the content of the plastic toy image block. By using mutually transposed convolution kernels, upsampling operations in convolutional neural networks can be implemented, that is, the size of the input matrix is ​​enlarged to obtain a higher resolution feature representation. This is very useful for tasks such as image generation and segmentation. When using convolution kernels for convolution operations, the spatial information of the image data can be retained. By using mutually transposed convolution kernels, the spatial structure of the original image block can be maintained while extracting features, which helps the network to better understand the image content. Mutually transposed convolution kernels can learn a reversible mapping from the input matrix to the output feature matrix, which is very useful for some tasks that require reconstruction of the original image block (such as image denoising, super-resolution, etc.). Therefore, by using the first convolutional neural network model with mutually transposed convolution kernels, the feature matrix can be effectively extracted from the plastic toy image block input matrix, retaining spatial information and providing better feature representation for subsequent tasks.

[0051] Correspondingly, the pooling subunit 1225 is used to expand the plastic toy image block feature matrix into the plastic toy image block feature vector according to row vectors or column vectors. It should be understood that expanding the image block into a vector by row or column can ensure that the dimension of the feature vector matches the input dimension of the classifier. Doing so can simplify the feature representation and processing process, making the feature vector easier to be processed by the classifier. Expanding the image block into a vector can better retain the feature information inside the image block. Each element represents a feature in the image block. Through this expansion method, the features of the image block can be more comprehensively described. Expanding the image block into a vector by row or column can make the feature vectors of different image blocks more independent. This helps the classifier to better distinguish the features of different image blocks and improve the accuracy of classification. Expanding the image block into a vector can simplify the calculation process and reduce the amount of calculation. This is very important for the processing and feature extraction of large-scale image data, and can improve the operating efficiency of the system. Most machine learning models (such as neural networks, support vector machines, etc.) usually accept input in the form of vectors. Therefore, expanding the image block into a vector by row or column can be directly used as the input of these models, which is convenient for further feature processing and classification.

[0052] In an embodiment of the present application, the plastic toy detection fusion module 130 is used to perform a topological modulation-based perspective consistency matching fusion on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector. It should be understood that fusing the plastic toy enhanced feature vector with the plastic toy image block feature vector can combine feature information of different levels and types, thereby more comprehensively describing the features of the plastic toy image. This helps to improve the classifier's understanding of the image and classification accuracy. The plastic toy enhanced feature vector may contain information after the image is enhanced, such as rotation, scaling, brightness adjustment, etc. Combining these enhanced features with the features of the original image block can make the model more robust and better able to cope with images under different conditions. By fusing features from different sources, the generalization ability of the model can be improved, so that it can better adapt to different plastic toy image samples, not just samples in the training set. Fusion of different types of features can help reduce the risk of overfitting of the model, because the model can learn feature information about plastic toy images from multiple perspectives, rather than relying solely on a single type of feature. By constructing a plastic toy image classification feature vector between the plastic toy enhancement feature vector and the plastic toy image block feature vector, the performance and robustness of the classifier can be improved, so that it can be better applied to the actual plastic toy image classification task.

[0053] In particular, in the technical solution of the present application, considering that the plastic toy enhanced feature vector and the plastic toy image block feature vector are extracted from different processing paths, they represent different types of image features. Due to their essential differences and different extraction methods, the two feature vectors have significant feature heterogeneity. Specifically, extracted from the plastic toy multi-channel image, processed by the grayscale co-occurrence matrix and the residual dual attention mechanism model, the global texture and structural features of the image are mainly focused on. Extracted from the plastic toy image block sequence, processed by the embedding layer and the convolutional neural network model, the local details and block features of the image are mainly focused on. When the plastic toy enhanced feature vector and the plastic toy image block feature vector are fused, the plastic toy enhanced feature vector and the plastic toy image block feature vector may exist on different manifolds, which means that they have different distribution laws in their respective spaces. Directly splicing them together may result in the final feature vector not having a clear manifold in the new space, that is, the feature manifold smoothness is insufficient. Since the two feature vectors come from different sources, they may not be well mapped to each other in the fused feature space, that is, one type of feature may not accurately reflect the feature changes of another type, resulting in increased uncertainty in feature expression. In order to solve this problem, in the technical solution of the present application, the plastic toy enhanced feature vector and the plastic toy image block feature vector are subjected to perspective consistency matching fusion based on topological modulation processing.

[0054] Specifically, in an embodiment of the present application, the plastic toy enhanced feature vector and the plastic toy image block feature vector are subjected to perspective consistency matching fusion based on topological modulation processing, including: constructing a plastic toy image distance topological matrix between the plastic toy enhanced feature vector and the plastic toy image block feature vector; constructing a plastic toy image eigenvalue granularity association matrix between the plastic toy enhanced feature vector and the plastic toy image block feature vector; performing topological association modulation on the plastic toy image eigenvalue granularity association matrix based on the plastic toy image distance topological matrix to obtain a plastic toy image topological modulation association matrix; using the plastic toy image topological modulation association matrix as a perspective co-projection space, projecting the plastic toy enhanced feature vector and the plastic toy image block feature vector to the perspective co-projection space to obtain a perspective modulated plastic toy enhanced feature vector and a perspective modulated plastic toy image block feature vector; and fusing the perspective modulated plastic toy enhanced feature vector and the perspective modulated plastic toy image block feature vector to obtain the plastic toy image classification feature vector.

[0055] In an embodiment of the present application, another implementation expression of performing perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector can be: performing perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector according to the following formula to obtain a plastic toy image classification feature vector; wherein the formula is:

[0056] M t =D(V1,V2)

[0057]

[0058] V c =α·V1 ′ +β·V2 ′

[0059] Wherein, V1 represents the plastic toy enhanced feature vector, V2 represents the plastic toy image block feature vector, and D(V1, V2) is the plastic toy image distance topological matrix between the plastic toy enhanced feature vector and the plastic toy image block feature vector, that is, d i,j =d(v 1i ,v 2j ), d i,j is the eigenvalue of the position (i, j) in the distance topological matrix of the plastic toy image, d(v 1i ,v 2j) represents the distance between the eigenvalue of the ith position of the plastic toy enhanced feature vector and the eigenvalue of the jth position of the plastic toy image block feature vector, and V1 and V2 are both column vectors, M t represents the distance topological matrix of the plastic toy image, T represents the transpose of the vector, Represents matrix multiplication, M v represents the eigenvalue granularity correlation matrix of the plastic toy image, exp represents the natural exponential function, M c represents the topological modulation correlation matrix of the plastic toy image, V1 ′ Represents the enhanced feature vector of the view-modulated plastic toy, V2 ′ represents the feature vector of the view-modulated plastic toy image patch, α and β represent weighted hyperparameters, V c Represents the feature vector for plastic toy image classification.

[0060] That is, in the technical solution of the present application, the plastic toy enhanced feature vector and the plastic toy image block feature vector are subjected to perspective consistency matching fusion based on topological modulation processing, and the plastic toy image distance topological matrix between the plastic toy enhanced feature vector and the plastic toy image block feature vector is first constructed. This step utilizes the idea of ​​graph theory, takes the eigenvalues ​​of each position in the plastic toy enhanced feature vector and the plastic toy image block feature vector as nodes, and uses the distance metric function to determine the low-dimensional embedding expression of the edges between nodes, so as to construct a feature expression that can quantify the topological association relationship between the plastic toy enhanced feature vector and the plastic toy image block feature vector. This step is crucial to understanding the spatial distribution of data points and their mutual connections, and also provides basic structural information for subsequent steps.

[0061] Next, a plastic toy image feature value granularity association matrix is ​​constructed between the plastic toy enhanced feature vector and the plastic toy image block feature vector. This step focuses on exploring the correlation between the internal attributes of the feature vector. In a specific example, the product between the plastic toy enhanced feature vector and the transposed vector of the plastic toy image block feature vector is calculated to obtain the plastic toy image feature value granularity association matrix. Furthermore, the plastic toy image feature value granularity association matrix is ​​topologically associated modulated based on the plastic toy image distance topological matrix to obtain a plastic toy image topological modulation association matrix. That is, the feature value granularity association information of the plastic toy enhanced feature vector and the plastic toy image block feature vector is driven to wander on the distance topological map to construct a low-dimensional modulated perspective co-projection space domain for mapping the plastic toy enhanced feature vector and the plastic toy image block feature vector to a continuous high-dimensional regression space attribute.

[0062] Next, the plastic toy image topological modulation association matrix is ​​used as the perspective co-projection space, and the plastic toy enhanced feature vector and the plastic toy image block feature vector are projected into the perspective co-projection space to obtain the perspective modulated plastic toy enhanced feature vector and the perspective modulated plastic toy image block feature vector. That is, the key to this technical solution is to find an implicit third space with modulation capability, in which as much information of the original high-dimensional data as possible is retained, so that the converted new feature vector not only reduces the dimension, but also maintains the relative position relationship between the original features.

[0063] Finally, the perspective modulated plastic toy enhancement feature vector and the perspective modulated plastic toy image block feature vector are fused to obtain the plastic toy image classification feature vector. Fusion can be achieved in a variety of forms, ranging from simple arithmetic average to complex ensemble learning algorithms, depending on the application scenario and the goals pursued. The final plastic toy image classification feature vector combines the main features of the two sets of input data and provides a more comprehensive data representation, which is not only conducive to improving the performance of machine learning models, but also promotes the ability of cross-modal or multi-source data analysis, thereby significantly enhancing the understanding and utilization efficiency of complex data structures.

[0064] In the embodiment of the present application, the plastic toy detection and analysis module 140 is used to pass the plastic toy image classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be detected is defective. It should be understood that in this case, our goal is to classify plastic toys, that is, to determine whether they are defective. By inputting the extracted feature vector into the classifier, we can use machine learning algorithms (such as support vector machines, logistic regression, neural networks, etc.) to train the model so that it can automatically identify and classify different categories of plastic toys. The output result of the classifier can be used as a decision basis to help determine whether the plastic toy to be detected is defective. According to the prediction results of the classifier, corresponding actions can be taken, such as removing defective plastic toys from the production line or repairing them. By using a classifier, automatic detection and classification of a large number of plastic toy images can be achieved, improving production efficiency and quality control levels. This automated detection system can quickly and accurately identify defects and avoid interference from human factors. Therefore, passing the plastic toy image classification feature vector through a classifier to obtain a classification result, and then using the result to indicate whether the plastic toy to be detected is defective, is an important step to achieve automated detection, improve production efficiency and quality control levels.

[0065] In summary, the Al-based plastic toy visual inspection system and method according to the embodiment of the present application acquires plastic toy images, extracts feature vectors and constructs classification feature vectors, and uses classifiers to determine whether the plastic toys have defects, thereby realizing automated quality inspection and classification identification, replacing traditional manual quality inspection methods, and improving efficiency and accuracy.

[0066] As described above, the Al-based plastic toy visual inspection system 100 according to the embodiment of the present application can be implemented in various terminal devices, such as a server of the Al-based plastic toy visual inspection system, etc. In one example, the Al-based plastic toy visual inspection system 100 can be integrated into the terminal device as a software module and / or a hardware module. For example, the Al-based plastic toy visual inspection system 100 can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the Al-based plastic toy visual inspection system 100 can also be one of the many hardware modules of the terminal device.

[0067] Alternatively, in another example, the Al-based plastic toy visual inspection system 100 and the terminal device may also be separate devices, and the Al-based plastic toy visual inspection system 100 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0068] Figure 5 FIG. 1 is a flow chart of a plastic toy visual inspection method based on Al according to an embodiment of the present application. Figure 5 As shown, according to the Al-based plastic toy visual inspection method of the embodiment of the present application, the method includes the following steps: S110, obtaining a plastic toy image of the plastic toy to be inspected; S120, extracting a plastic toy enhanced feature vector and a plastic toy image block feature vector from the plastic toy image of the plastic toy to be inspected respectively; S130, performing perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector; S140, passing the plastic toy image classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be inspected has defects.

[0069] Here, those skilled in the art will appreciate that the specific operations of each step in the above-mentioned plastic toy visual inspection method based on Al have been described in the above reference. Figures 1 to 4 It has been introduced in detail in the description of the Al-based plastic toy visual inspection system, and therefore, its repeated description will be omitted.

[0070] The basic principles of the present disclosure are described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, effects, etc. mentioned in the present disclosure are only examples and not limitations, and it cannot be considered that these advantages, strengths, effects, etc. are required by each embodiment of the present disclosure. In addition, the specific details disclosed above are only for the purpose of illustration and ease of understanding, and are not limitations. The above details do not limit the present disclosure to the necessity of adopting the above specific details to be implemented.

[0071] The block diagrams of the devices, apparatuses, equipment, and systems involved in this disclosure are only illustrative examples and are not intended to require or imply that they must be connected, arranged, and configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems can be connected, arranged, and configured in any manner. Words such as "including," "comprising," "having," and the like are open words, referring to "including but not limited to," and can be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or," and can be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and can be used interchangeably therewith.

[0072] Additionally, as used herein, "or" used in a list of items beginning with "at least one" indicates a separate list, so that, for example, a list of "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not mean that the example described is preferred or better than other examples.

[0073] It should also be noted that in the system and method of the present disclosure, each component or each step can be decomposed and / or recombined. Such decomposition and / or recombination should be regarded as equivalent solutions of the present disclosure.

[0074] Various changes, substitutions, and modifications of the techniques described herein may be made without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of the present disclosure is not limited to the specific aspects of the processes, machines, manufactures, compositions of events, means, methods, and actions described above. Currently existing or later to be developed processes, machines, manufactures, compositions of events, means, methods, or actions that perform substantially the same functions or achieve substantially the same results as the corresponding aspects described herein may be utilized. Thus, the appended claims include such processes, machines, manufactures, compositions of events, means, methods, or actions within their scope.

[0075] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present disclosure. Therefore, the present disclosure is not intended to be limited to the aspects shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

[0076] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present disclosure to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A plastic toy visual inspection system based on Al, characterized in that: include: A plastic toy detection and acquisition module, used to obtain a plastic toy image of a plastic toy to be detected; A plastic toy detection processing module, used to extract a plastic toy enhancement feature vector and a plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected; A plastic toy detection fusion module, used for performing view consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector; The plastic toy detection and analysis module is used to pass the plastic toy image classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be detected has defects.

2. The Al-based plastic toy visual inspection system according to claim 1 is characterized in that: The plastic toy detection and processing module comprises: A texture feature extraction unit, used for extracting texture features from the plastic toy image of the plastic toy to be detected to obtain the plastic toy enhanced feature vector; The block feature extraction unit is used to extract the plastic toy image block feature vector from the plastic toy image of the plastic toy to be detected.

3. The Al-based plastic toy visual inspection system according to claim 2 is characterized in that: The texture feature extraction unit comprises: A gray level co-occurrence matrix subunit, used for extracting a plastic toy texture feature map from the plastic toy image of the plastic toy to be detected based on a gray level co-occurrence matrix; A cascade subunit, used for cascading the plastic toy image and the plastic toy texture feature map to obtain a plastic toy multi-channel image; A residual dual attention subunit, used for passing the plastic toy multi-channel image through a residual dual attention mechanism model to obtain a plastic toy enhanced multi-channel feature map; The plastic toy feature dimensionality reduction subunit is used to perform global mean pooling on each feature matrix along the channel dimension of the plastic toy enhanced multi-channel feature map to obtain the plastic toy enhanced feature vector.

4. The Al-based plastic toy visual inspection system according to claim 3 is characterized in that: The residual dual attention sub-unit comprises: A spatial attention secondary subunit, used for passing the multi-channel image of the plastic toy through the spatial attention module of the residual dual attention mechanism model to obtain a plastic toy spatial attention map; A channel attention secondary subunit, used for passing the plastic toy multi-channel image through the channel attention module of the residual dual attention mechanism model to obtain a plastic toy channel attention map; A fusion weighted secondary subunit is used to calculate the position points between the plastic toy spatial attention map and the plastic toy channel attention map to obtain a plastic toy weighted feature map; The fusion feature secondary subunit is used to fuse the plastic toy multi-channel image and the plastic toy weighted feature map to obtain the plastic toy enhanced feature map.

5. The Al-based plastic toy visual inspection system according to claim 4 is characterized in that: The spatial attention secondary subunit comprises: A pooling three-level sub-unit, used for performing average pooling and maximum pooling on the plastic toy multi-channel image along the channel dimension to obtain a plastic toy average feature matrix and a plastic toy maximum feature matrix; A cascade three-stage sub-unit, used for cascading and channel-adjusting the plastic toy average characteristic matrix and the plastic toy maximum characteristic matrix to obtain a plastic toy channel characteristic matrix; A convolutional coding tertiary subunit, used for performing convolutional coding on the plastic toy channel feature matrix using the convolutional layer of the spatial attention module to obtain a plastic toy convolutional feature matrix; The third-level subunit of the Softmax function is used to pass the plastic toy convolution feature matrix through the Softmax function to obtain a plastic toy spatial attention score matrix; The three-level sub-unit of position point multiplication is used to perform position point multiplication on the plastic toy spatial attention score matrix and each feature matrix of the plastic toy multi-channel image along the channel dimension to obtain the plastic toy spatial attention map.

6. The Al-based plastic toy visual inspection system according to claim 5, characterized in that: The block feature extraction unit comprises: An image block processing subunit, used for dividing the plastic toy image into image blocks to obtain a plastic toy image block sequence; A linear encoding subunit, used for encoding the plastic toy image block sequence through a linear embedding layer to obtain a plurality of plastic toy image block embedding vectors; A two-dimensional arrangement subunit, used for two-dimensionally arranging the plurality of plastic toy image block embedding vectors to obtain a plastic toy image block input matrix; A convolutional coding subunit, used for inputting the plastic toy image block input matrix through adjacent layers using a first convolutional neural network model with mutually transposed convolution kernels to obtain a plastic toy image block feature matrix; The pooling subunit is used to expand the plastic toy image block feature matrix into the plastic toy image block feature vector according to row vectors or column vectors.

7. The Al-based plastic toy visual inspection system according to claim 6, characterized in that: The plastic toy detection fusion module is used to: Constructing a plastic toy image distance topology matrix between the plastic toy enhancement feature vector and the plastic toy image block feature vector; Constructing a plastic toy image feature value granularity correlation matrix between the plastic toy enhancement feature vector and the plastic toy image block feature vector; Performing topological correlation modulation on the plastic toy image eigenvalue granularity correlation matrix based on the plastic toy image distance topological matrix to obtain a plastic toy image topological modulation correlation matrix; Taking the plastic toy image topological modulation association matrix as the perspective co-projection space, projecting the plastic toy enhancement feature vector and the plastic toy image block feature vector into the perspective co-projection space to obtain the perspective modulated plastic toy enhancement feature vector and the perspective modulated plastic toy image block feature vector; The view-modulated plastic toy enhancement feature vector and the view-modulated plastic toy image block feature vector are fused to obtain the plastic toy image classification feature vector.

8. A plastic toy visual inspection method based on Al, characterized in that: include: Acquire a plastic toy image of the plastic toy to be detected; Extracting a plastic toy enhanced feature vector and a plastic toy image block feature vector from the plastic toy image to be detected; Performing perspective consistency matching fusion based on topological modulation processing on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector; The classification feature vector of the plastic toy image is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the plastic toy to be inspected has defects.

9. The Al-based plastic toy visual inspection method according to claim 8, characterized in that: Extracting a plastic toy enhancement feature vector and a plastic toy image block feature vector from the plastic toy image to be detected respectively includes: Extracting texture features from the plastic toy image of the plastic toy to be detected to obtain the plastic toy enhanced feature vector; The plastic toy image block feature vector is extracted from the plastic toy image of the plastic toy to be detected.

10. The Al-based plastic toy visual inspection method according to claim 9, characterized in that: Performing a topological modulation-based perspective consistency matching fusion on the plastic toy enhanced feature vector and the plastic toy image block feature vector to obtain a plastic toy image classification feature vector, including: Constructing a plastic toy image distance topology matrix between the plastic toy enhancement feature vector and the plastic toy image block feature vector; Constructing a plastic toy image feature value granularity correlation matrix between the plastic toy enhancement feature vector and the plastic toy image block feature vector; Performing topological correlation modulation on the plastic toy image eigenvalue granularity correlation matrix based on the plastic toy image distance topological matrix to obtain a plastic toy image topological modulation correlation matrix; Taking the plastic toy image topological modulation association matrix as the perspective co-projection space, projecting the plastic toy enhancement feature vector and the plastic toy image block feature vector into the perspective co-projection space to obtain the perspective modulated plastic toy enhancement feature vector and the perspective modulated plastic toy image block feature vector; The view-modulated plastic toy enhancement feature vector and the view-modulated plastic toy image block feature vector are fused to obtain the plastic toy image classification feature vector.

Images