Enamel crack sorting method, device, computer equipment and storage medium

Abstract

The embodiment of the present invention discloses an enamel crack sorting method, device, computer equipment and storage medium, which belongs to the field of image processing technology. The embodiment of the present invention uses a set of scanned images of the surface of enamel products and adopts wavelet descriptor recognition technology to analyze the surface cracks. First, the initial crack wavelet descriptor is extracted to lay the foundation for subsequent analysis. Based on this, the scanned images are summarized to obtain image batches and corresponding algorithm indexes, which can improve processing efficiency and adapt different batch algorithms. Multidimensional wavelet descriptor recognition is performed on each batch according to the index to obtain the target edge morphological texture vector, and hierarchical and targeted processing can analyze the surface features in detail. The decision wavelet descriptor is obtained by splicing the vectors of each batch, and then the crack sorting label is obtained. The whole process comprehensively and accurately sorts cracks, improves the efficiency and accuracy of enamel product quality inspection, reduces missed detection losses, and provides a reliable basis for related work.

Description

Technical Field

[0001] The embodiments of the present invention belong to the field of image processing technology, and specifically relate to an enamel crack sorting method, device, computer equipment and storage medium. Background Art

[0002] During the production, use and preservation of enamel products, cracks may appear on the surface. Accurate detection and sorting of these cracks are of vital importance for the quality control, repair and preservation of enamel products. However, there are many technical problems with traditional crack detection methods for enamel products. For example, traditional detection methods often lack a systematic and comprehensive detection process, and may simply scan and observe the surface, making it difficult to accurately identify the characteristics of the cracks. This inaccuracy makes it easy to miss some minor cracks during the quality inspection process, thus affecting the quality of enamel products. Moreover, due to the lack of effective classification and processing methods, it is difficult to conduct targeted analysis of cracks with different characteristics, resulting in low detection efficiency. Summary of the invention

[0003] The embodiments of the present invention provide an enamel crack sorting method, device, computer equipment and storage medium, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] An embodiment of the present invention provides an enamel crack sorting method, which is applied to a computer device. The method comprises: obtaining a target scan image set consisting of a plurality of enamel product surface scan images; performing wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor; and summarizing the enamel product surface scan images in the target scan image set based on the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, wherein each of the surface scan image batches includes a plurality of the enamel product surface scan images, and the linkage image algorithm indexes An index corresponding to an image description processing algorithm for implementing multidimensional wavelet descriptor recognition; for each batch of surface scan images, multidimensional wavelet descriptor recognition is performed on the initial crack wavelet descriptor according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch, so as to obtain a target edge morphological texture vector corresponding to the surface scan image batch; the target edge morphological texture vectors corresponding to each batch of surface scan images are concatenated to obtain a crack sorting decision wavelet descriptor corresponding to the target scan image set; the crack sorting decision wavelet descriptor is processed to obtain an enamel crack sorting label.

[0005] An embodiment of the present invention provides an enamel crack sorting device, comprising:

[0006] An image recognition module is used to obtain a target scan image set consisting of a plurality of scan images of the surface of enamel products; perform wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor;

[0007] An image summarization module is used to summarize the surface scan images of the enamel products in the target scan image set according to the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, each of the surface scan image batches including a plurality of surface scan images of the enamel products, and the linkage image algorithm index is an index corresponding to an image description processing algorithm for implementing multidimensional wavelet descriptor recognition;

[0008] an algorithm matching module, for performing multidimensional wavelet descriptor recognition on the initial crack wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scanning image batch for each of the surface scanning image batches, to obtain a target edge morphology texture vector corresponding to the surface scanning image batch;

[0009] The crack sorting module is used to splice the target edge morphology texture vectors corresponding to each batch of the surface scanned images to obtain the crack sorting decision wavelet descriptor corresponding to the target scanned image set; and process the crack sorting decision wavelet descriptor to obtain the enamel crack sorting label.

[0010] An embodiment of the present invention provides a computer device, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0011] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0012] The embodiment of the present invention obtains a set of scanned images of the surface of enamel products and adopts wavelet descriptor recognition technology to analyze the surface cracks. First, the initial crack wavelet descriptor is extracted to lay the foundation for subsequent analysis. Based on this, the scanned images are summarized to obtain image batches and corresponding algorithm indexes, which can improve processing efficiency and adapt different batches of algorithms. Multi-dimensional wavelet descriptor recognition is performed on each batch according to the index to obtain the target edge morphological texture vector. Hierarchical and targeted processing can analyze the surface features in detail. The decision wavelet descriptor is obtained by splicing the vectors of each batch, and then the crack sorting label is obtained. The whole process comprehensively and accurately sorts cracks, improves the efficiency and accuracy of enamel product quality inspection, reduces missed detection losses, and provides a reliable basis for related work. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention provides a flowchart of an enamel crack sorting method according to an embodiment of the present invention.

[0014] Figure 2 A schematic diagram of the structure of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments in the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present invention.

[0016] The terms "first", "second", etc. in the embodiments of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the embodiments of the present invention can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the embodiments of the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0017] Figure 1 A method for sorting enamel cracks is shown, which is applied to computer equipment. The method includes the following S110-S140.

[0018] S110, the computer device obtains a target scan image set consisting of a plurality of enamel product surface scan images; performs wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor.

[0019] The computer equipment first obtains a number of enamel product surface scan images through a high-resolution scanning device, and the enamel product surface scan images together constitute a target scan image set. Subsequently, the computer equipment uses wavelet analysis technology to perform wavelet descriptor recognition on each enamel product surface scan image in the target scan image set to extract the initial crack wavelet descriptor that characterizes the crack characteristics.

[0020] In detail, as far as image acquisition is concerned, the computer equipment performs a full-scale, high-resolution scan of the surface of the enamel product by connecting to a high-precision scanning device, such as a professional-grade digital camera or a laser scanner, to ensure that the acquired image can clearly show all the details of the product surface, including tiny cracks.

[0021] As far as wavelet descriptor identification is concerned, wavelet analysis can capture multi-scale features in images. In this step, the computer equipment uses wavelet transform to decompose the scanned image, extract the wavelet coefficients containing crack information, and then form the initial crack wavelet descriptor. These initial crack wavelet descriptors not only contain the location information of the crack, but also contain multi-dimensional features such as the shape, length, and width of the crack, providing a rich data basis for subsequent analysis.

[0022] S120. The computer device summarizes the surface scan images of the enamel products in the target scan image set based on the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, wherein each of the surface scan image batches includes a plurality of surface scan images of the enamel products, and the linkage image algorithm index is an index corresponding to an image description processing algorithm that implements multidimensional wavelet descriptor recognition.

[0023] Based on the initial crack wavelet descriptor, the computer device classifies the images in the target scan image set to form several surface scan image batches, and assigns a linkage image algorithm index to each surface scan image batch. The linkage image algorithm index points to a corresponding multidimensional wavelet descriptor recognition algorithm for further in-depth analysis of the crack characteristics of the batch of images.

[0024] In detail, the computer device groups the scanned images according to the similarity of the initial crack wavelet descriptors using a clustering algorithm (such as K-means or DBSCAN). Images with high similarity are grouped into the same batch, which helps to conduct a more refined analysis of specific types of cracks later.

[0025] The computer equipment presets a variety of multidimensional wavelet descriptor recognition algorithms for different types of cracks. These multidimensional wavelet descriptor recognition algorithms have their own advantages in processing complex crack morphology and edge details. The computer equipment automatically selects the most suitable algorithm based on the crack characteristics of each batch and generates the corresponding linkage image algorithm index. This step ensures the pertinence and accuracy of subsequent analysis.

[0026] S130. The computer device performs multidimensional wavelet descriptor recognition on the initial crack wavelet descriptor for each batch of surface scan images according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the batch of surface scan images, and obtains the target edge morphology texture vector corresponding to the batch of surface scan images.

[0027] For each batch of surface scanned images, the computer device calls the corresponding image description processing algorithm according to its corresponding linkage image algorithm index to perform multi-dimensional wavelet descriptor recognition on the initial crack wavelet descriptor. This process aims to deeply explore the edge morphology and texture features of the cracks and finally generate the target edge morphology texture vector corresponding to the batch.

[0028] Compared with the basic wavelet descriptor in S110, the multidimensional wavelet descriptor provides a richer and more detailed feature representation. The computer device further decomposes and reorganizes the initial crack wavelet descriptor through the algorithm indicated by the linkage image algorithm index, and extracts higher-dimensional crack features, such as the curvature, branching structure, and local texture changes of the crack. After identification by the multidimensional wavelet descriptor, the computer device integrates these high-dimensional features into the target edge morphology texture vector. The target edge morphology texture vector fully reflects the morphology and texture characteristics of the crack in numerical form, providing solid data support for subsequent crack classification and decision-making.

[0029] S140, the computer equipment splices the target edge morphology texture vectors corresponding to each batch of the surface scan images to obtain the crack sorting decision wavelet descriptor corresponding to the target scan image set; processes the crack sorting decision wavelet descriptor to obtain the enamel crack sorting label.

[0030] The computer equipment splices the target edge morphology texture vectors corresponding to each surface scan image batch to construct a crack sorting decision wavelet descriptor corresponding to the target scan image set. Subsequently, the computer equipment performs comprehensive analysis and processing on the descriptor to finally generate the sorting label of the enamel crack.

[0031] In detail, the computer equipment forms a comprehensive descriptor (crack sorting decision wavelet descriptor) containing the crack characteristics of all batches by splicing the target edge morphology texture vectors of each batch. The crack sorting decision wavelet descriptor not only covers the local characteristics of the cracks, but also reflects the overall distribution law and mutual relationship of the cracks.

[0032] Furthermore, based on the crack sorting decision wavelet descriptor, the computer equipment uses a machine learning algorithm (such as a support vector machine, a random forest or a deep learning model) to classify the cracks. During the classification process, different types of cracks, such as linear cracks, reticular cracks, and cracks, can be automatically learned and identified based on the features in the crack sorting decision wavelet descriptor, and a unique sorting label can be assigned to each crack. These labels are of great guiding significance for the subsequent protection, restoration and research of enamel products.

[0033] To facilitate a clearer implementation of the above S110 - S140 , a complete example is used for introduction below.

[0034] The computer equipment first obtains the scanned images of the enamel product surface through a high-resolution scanning device, and these images together constitute the target scanned image set. Subsequently, the wavelet analysis technology is used to identify the wavelet descriptor of each image in the set to extract the initial crack wavelet descriptor that characterizes the crack characteristics.

[0035] For example, there is a collection of 50 enamel products. The surface of each product is scanned by a professional digital camera with a resolution of 600dpi to ensure the clarity and detail capture of the image. During the scanning process, uniform lighting and multiple angles are used to avoid the impact of shadows and reflections on image quality. Finally, a 512*512 pixel grayscale image is obtained for each enamel product.

[0036] Furthermore, Db2 wavelet is selected as the transform basis because it has good time-frequency localization characteristics and is suitable for capturing edge and texture features in images. Three-level wavelet decomposition is performed on each scanned image to obtain crack information at different scales. During the decomposition process, the image is decomposed into a low-frequency part (approximate coefficient) and a high-frequency part (detail coefficient), where the high-frequency part contains edge information such as cracks. The wavelet coefficients of the high-frequency part are extracted, and after threshold processing, the initial crack wavelet descriptor is formed. The purpose of threshold setting is to remove noise interference and retain effective crack information. For example, for a 512*512 pixel image, after wavelet transform and threshold processing, a descriptor vector containing thousands of wavelet coefficients may be obtained. These coefficients reflect the distribution of cracks at different scales and positions.

[0037] Take a scanned image of an enamel product with cracks as an example. After wavelet transform, a wavelet coefficient matrix containing three-level decomposition is obtained. In this example, only the coefficients of the high-frequency part are concerned, so for each scale, a coefficient matrix of the same size as the original image can be obtained. By splicing these coefficient matrices according to scale, an initial crack wavelet descriptor vector can be obtained. For example, for a three-level decomposition, a descriptor vector with a length of 512*512*3 (or more precisely, the total number of coefficients after removing the low-frequency part) may be obtained.

[0038] Secondly, based on the initial crack wavelet descriptor, the computer device classifies the images in the target scan image set to form several surface scan image batches, and assigns a linkage image algorithm index to each batch. The index points to a corresponding multidimensional wavelet descriptor recognition algorithm for further in-depth analysis of the crack characteristics of the batch of images.

[0039] In detail, the K-means clustering algorithm can be used to group the initial crack wavelet descriptors. The number of clusters is set to 5, that is, the images are expected to be divided into 5 batches. During the clustering process, the Euclidean distance is used as the similarity metric, and the number of iterations is set to 100 times to ensure the stability of the clustering results. Finally, 5 image batches containing similar crack features are obtained.

[0040] Linked image algorithm index: Five multidimensional wavelet descriptor recognition algorithms are preset, each optimized for different types of crack features. For example, algorithm 1 may be good at processing linear cracks, algorithm 2 may be good at processing mesh cracks, etc. According to the crack characteristics of each batch, the most suitable algorithm is automatically selected and the corresponding linked image algorithm index is generated. For example, for a batch containing a large number of linear cracks, the computer device may select algorithm 1 and generate an index pointing to algorithm 1.

[0041] For another example, there is a collection of 100 scanned images, which are divided into 5 batches after K-means clustering, each batch containing 20 images. For the first batch, the computer device selects algorithm 1 for multidimensional wavelet descriptor recognition and generates an index value (such as 1) for it. For the second batch, the computer device selects algorithm 2 and generates another index value (such as 2) for it. And so on, until all batches are assigned corresponding algorithms and indexes.

[0042] Then, for each batch of surface scanned images, the computer device calls the corresponding image description processing algorithm according to its corresponding linkage image algorithm index to perform multi-dimensional wavelet descriptor recognition on the initial crack wavelet descriptor. This process aims to deeply explore the edge morphology and texture features of the cracks and finally generate the target edge morphology texture vector corresponding to the batch.

[0043] As for multidimensional wavelet descriptor identification, taking Db4 wavelet as an example (different wavelets are selected in this embodiment to show diversity), five-level wavelet decomposition is performed to obtain richer crack features. During the decomposition process, the coefficients of the high-frequency part are also focused because they contain the edge and texture information of the crack.

[0044] After extracting the wavelet coefficients of the high-frequency part, their features such as histogram of oriented gradients (HOG) or local binary pattern (LBP) can be further calculated to capture the edge morphology and texture information of the crack. These features can reflect the complex characteristics of the crack, such as curvature, branching structure, and local texture changes.

[0045] For each image in each batch, a vector containing multidimensional wavelet descriptor features can be obtained. After normalization (such as scaling the eigenvalues ​​to the range of [0, 1]), these vectors can be used for subsequent crack classification and decision making.

[0046] For example, a multidimensional wavelet descriptor recognition algorithm is selected for each batch, and a feature vector of length N is generated for each image. Then for a batch containing M images, an M*N feature matrix can be obtained. This matrix is ​​the target edge morphology texture vector set corresponding to the batch.

[0047] Taking the first batch as an example, Algorithm 1 is selected for multidimensional wavelet descriptor recognition, and each image generates a feature vector of length 100. Then for a batch of 20 images, a 20*100 feature matrix can be obtained. Each row in this matrix represents the feature vector of an image, and each column represents the distribution of a feature on different images (due to space constraints, the elements in the feature matrix are not listed one by one).

[0048] Finally, the computer equipment splices the target edge morphology texture vectors corresponding to each surface scan image batch to construct a crack sorting decision wavelet descriptor corresponding to the target scan image set. Subsequently, the descriptor is comprehensively analyzed and processed to finally generate the sorting label of the enamel crack.

[0049] For example, there are 5 batches, each batch corresponds to a target edge morphology texture vector set containing 20 images (i.e., a 20*N feature matrix). By concatenating the row vectors (or column vectors, depending on the needs of subsequent processing) of these matrices, a comprehensive descriptor matrix containing crack features of all batches can be obtained. This matrix is ​​the crack sorting decision wavelet descriptor.

[0050] As for the generation of enamel crack sorting labels, support vector machine (SVM) can be used as a classifier to classify crack sorting decision wavelet descriptors. During the training process, crack samples of known types are used as training data to learn different types of crack characteristics. During the classification process, SVM will automatically learn and identify different types of cracks based on the features in the crack sorting decision wavelet descriptor. For example, it may identify linear cracks, network cracks, cracks and other types, and assign a unique sorting label to each crack (such as 1 for linear cracks, 2 for network cracks, etc.). Finally, a result set containing all image crack sorting labels can be obtained, which can be used for subsequent protection, restoration and research of enamel products.

[0051] For another example, there is a set of 100 scanned images. After the previous processing, a crack sorting decision wavelet descriptor matrix is ​​obtained (for example, 100*M, where M is the total number of all features). Then, this matrix is ​​classified using SVM, and a result set containing 100 sorting labels is obtained. For example, the first label in the result set may be 1, indicating that the crack in the first image is a linear crack; the second label may be 2, indicating that the crack in the second image is a mesh crack; and so on.

[0052] In addition, in order to evaluate the performance of computer equipment, a variety of indicators can be used for measurement, such as recognition accuracy, recall rate, F1 score, etc. At the same time, the advantages of computer equipment in recognition efficiency and accuracy can be evaluated by comparing with traditional manual recognition methods. In addition, methods such as cross-validation can be used to verify the stability and generalization ability of computer equipment.

[0053] In S140, when it comes to splicing each surface scan image batch with its corresponding target edge morphology texture vector, for example, there is a batch of 100 surface scan images, each image is preprocessed, and the edge morphology and texture feature vectors are extracted by edge detection algorithm (such as Canny edge detector) and texture analysis algorithm (such as gray level co-occurrence matrix GLCM). These vectors can contain different dimensions, for example, the edge morphology vector may be 128 dimensions, and the texture feature vector may be 64 dimensions. The splicing operation is to combine these vectors into a new, higher dimensional vector according to certain rules (such as direct connection or weighted fusion), for example, the dimension of the spliced ​​vector is 192 dimensions (128+64). In this way, for a batch of 100 images, a 100*192 matrix will be obtained, which is the preliminary form of the crack sorting decision wavelet descriptor.

[0054] Next, specific algorithm examples can be introduced for the processing stage of the crack sorting decision wavelet descriptor. For example, machine learning algorithms such as support vector machine (SVM) or random forest (RandomForest) can be used for classification. Before that, feature selection or dimensionality reduction operations can be performed to reduce computational complexity and improve classification performance. An exemplary dimensionality reduction method is principal component analysis (PCA), which can select the most representative principal components from a 192-dimensional feature vector. For example, 90% of the information is retained after dimensionality reduction, and the feature dimension may be reduced to about 30 dimensions. Then, the reduced feature vector is input into the SVM or random forest model and trained in combination with the known crack type labels. During the training and debugging process, the model will learn the mapping relationship between different crack types and feature vectors.

[0055] Finally, after the model training and debugging is completed, it can be used to classify the new crack sorting decision wavelet descriptor to obtain the enamel crack sorting label. For example, for a new batch of scanned images, its edge morphology and texture feature vectors are also extracted for splicing and dimensionality reduction, and then the processed feature vectors are input into the trained model. The model will output a probability distribution or a direct classification result, such as "crack type A, probability is 90%", which is the enamel crack sorting label.

[0056] In summary, the embodiment of the present invention can accurately analyze and process cracks that may exist on the surface of enamel products by acquiring a target scan image set of scanned images of the surface of enamel products and applying wavelet descriptor recognition technology. First, wavelet descriptor recognition can effectively extract the initial crack wavelet descriptor from the target scan image set, which is the basis for subsequent precise analysis. Then, the scanned images are summarized according to the initial crack wavelet descriptor to obtain different batches of surface scan images and corresponding linkage image algorithm indexes. This summarization method helps to improve processing efficiency, and corresponding image description processing algorithms are used according to the characteristics of different batches. Multidimensional wavelet descriptor recognition is performed for each batch based on its linkage image algorithm index to obtain the target edge morphology texture vector. This hierarchical and targeted processing method can analyze the characteristics of the surface of enamel products in more detail. The target edge morphology texture vectors corresponding to each batch are spliced ​​to obtain the crack sorting decision wavelet descriptor, and finally processed to obtain the enamel crack sorting label. The whole process realizes the comprehensive, systematic and accurate sorting of cracks in enamel products, which helps to improve the quality inspection efficiency and accuracy of enamel products and reduce the losses caused by inaccurate crack detection. It also provides a reliable basis for the production, repair and preservation of enamel products.

[0057] In some preferred embodiments, the surface scan images of enamel products in the target scan image set are summarized based on the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, including:

[0058] (1) performing attention feature embedding on the initial crack wavelet descriptor to obtain a crack attention embedding vector;

[0059] (2) identifying the initial screening batch category weight corresponding to each of the enamel product surface scan images according to the crack attention embedding vector;

[0060] (3) performing preliminary screening on the category weights of each of the preliminary screening batches to determine at least one candidate enamel product surface scan image, wherein the candidate enamel product surface scan image is an enamel product surface scan image whose corresponding category weight meets the set conditions;

[0061] (4) Obtaining the plurality of surface scan image batches based on the surface scan images of the candidate enamel products.

[0062] In practical applications, the specific technical solution of summarizing the surface scan images of enamel products in the target scan image set according to the initial crack wavelet descriptor to obtain several surface scan image batches and the linkage image algorithm index corresponding to each batch will be described in detail below.

[0063] For example, the initial crack wavelet descriptor is a feature vector with a dimension of n, such as n=1000. This vector contains multi-dimensional information of the crack, such as the location, shape, length, width and other information of the crack. The attention feature embedding process can adopt an attention mechanism based on a neural network, such as a multi-head attention mechanism. The multi-head attention mechanism can learn the feature relationship of the initial crack wavelet descriptor from different representation subspaces. By inputting the initial crack wavelet descriptor into the multi-head attention network, after a series of linear transformations, scaled dot product attention calculations and other operations, the crack attention embedding vector is finally obtained. The dimension of this vector may be m, for example, m=800. While retaining the key information of the initial crack wavelet descriptor, it further highlights the feature parts that are important for classification.

[0064] Take the Convolutional Neural Network (CNN) as an example to identify the initial screening batch category weights. The crack attention embedding vector is used as the input of CNN, and the convolutional layer of CNN can automatically learn the different feature patterns in the vector. For example, there are k possible initial screening batch categories, such as k=5. After the forward propagation process of CNN, a vector with a dimension of k is output in the fully connected layer. Each element in this vector represents the initial screening batch category weight corresponding to the surface scan image of the enamel product. For example, for a certain surface scan image of an enamel product, the initial screening batch category weight vector obtained is [0.1, 0.3, 0.2, 0.15, 0.25], corresponding to 5 possible initial screening batch categories.

[0065] The initial screening condition is set as the category weight greater than 0.2. For the image in the above example, its weights in the second and fifth initial screening batch categories are 0.3 and 0.25 respectively, which meets the set conditions, so the image will be determined as a candidate enamel product surface scanning image. In the entire target scan image set, by performing such a preliminary screening process on each image, a batch of candidate enamel product surface scanning images that meet the conditions can be screened out.

[0066] For these candidate enamel product surface scan images, clustering algorithms (such as DBSCAN algorithm) can be used to divide the surface scan image batches. The DBSCAN algorithm is based on the density of data points. If the density of data points within a certain radius (for example, the radius is set to r=5) exceeds a certain threshold (for example, the density threshold is minPts=3), these data points are divided into a cluster, that is, a surface scan image batch. After clustering, the images in each batch have a high similarity. Then, according to the characteristics of each batch, for example, if the image crack features in a batch show more linear features, a multidimensional wavelet descriptor recognition algorithm that is good at processing linear cracks can be assigned to this batch as a linkage image algorithm index; if the image crack features in a batch show more mesh features, a corresponding algorithm index that is good at processing mesh cracks is assigned to it. In this way, the induction from the target scan image set to several surface scan image batches is completed, and the corresponding linkage image algorithm index is determined for each batch, which is helpful for subsequent more targeted analysis and processing of enamel product surface cracks.

[0067] In the next step, the surface scan images of enamel products in the target scan image set are summarized based on the initial crack wavelet descriptor to obtain a number of surface scan image batches and the linked image algorithm index corresponding to each of the surface scan image batches, and also includes: for each of the surface scan image batches, determining the corresponding image batch cross description vector in the crack attention embedding vector; determining a target algorithm matching heat map based on the image batch cross description vector, the target algorithm matching heat map indicating the matching processing coefficient corresponding to each image description processing algorithm; and determining the linked image algorithm index corresponding to the surface scan image batch based on the target algorithm matching heat map.

[0068] In a related embodiment of the embodiment of the present invention, for the process of summarizing the surface scan images of enamel products in the target scan image set based on the initial crack wavelet descriptor to obtain several surface scan image batches and the linked image algorithm index corresponding to each batch, there is the following supplementary implementation method.

[0069] First of all, it is necessary to make it clear that the crack attention embedding vector is a vector containing rich information. For example, this crack attention embedding vector is a vector with 1000 elements, and its element values ​​represent the results of various features related to the surface cracks of enamel products after being processed by the attention mechanism.

[0070] For each batch of surface scanned images, a cross-image batch description vector that can accurately describe the characteristics of the batch is determined from the crack attention embedding vector. Taking one of the surface scanned image batches as an example, a method based on feature importance evaluation is used to determine the cross-image batch description vector.

[0071] First, the feature importance of each element in the crack attention embedding vector is evaluated. For example, the features corresponding to the 1st element to the 100th element in the vector are evaluated, and their importance values ​​for the current surface scanning image batch are respectively from 0.01 to 1.00 (this embodiment is just a numerical range set for the convenience of example). Then, the features corresponding to several elements with higher importance values ​​are selected to form an image batch cross description vector. For example, select elements with importance values ​​greater than 0.5, such as these elements are the features corresponding to the 30th, 40th, 55th, 60th, and 70th elements, then the image batch cross description vector of this surface scanning image batch is a new vector composed of these elements in the crack attention embedding vector, for example, this new vector is [0.3, 0.4, 0.6, 0.5, 0.7] (the numerical values ​​in this embodiment are only examples, and are actually the values ​​of the corresponding elements in the crack attention embedding vector). In the same way, the corresponding image batch cross description vectors can be determined for other surface scanning image batches.

[0072] Next, we need to determine the target algorithm matching heat map based on the cross description vector of the image batch obtained above. For example, there are 5 image description processing algorithms, namely Algorithm 1, Algorithm 2, Algorithm 3, Algorithm 4 and Algorithm 5.

[0073] For each algorithm and each image batch cross description vector of each surface scan image batch, a matching processing coefficient must be determined. Taking Algorithm 1 and the surface scan image batch mentioned above as an example, a vector similarity calculation method is used to determine the matching processing coefficient. Algorithm 1 corresponds to a preset feature template vector, for example, this vector is [0.2, 0.3, 0.5, 0.4, 0.6]. Then calculate the similarity between this preset vector and the image batch cross description vector [0.3, 0.4, 0.6, 0.5, 0.7] of the surface scan image batch. This embodiment uses the reciprocal of the exemplary Euclidean distance to represent the similarity (just an example method). First, calculate the Euclidean distance, which is sqrt((0.3-0.2)^2+(0.4-0.3)^2+(0.6-0.5)^2+(0.5-0.4)^2+(0.7-0.6)^2). We get a value, for example, 0.14. Then its reciprocal 1 / 0.14 is approximately equal to 7.14, which is the matching processing coefficient of Algorithm 1 for this batch of surface scan images.

[0074] In the same way, the matching processing coefficients of Algorithm 2, Algorithm 3, Algorithm 4 and Algorithm 5 for this surface scan image batch can be calculated, for example, 3.5, 5.0, 8.0 and 2.0 are obtained respectively. In this way, a set of matching processing coefficients corresponding to this surface scan image batch and these five algorithms are obtained.

[0075] Repeat the above operation for other surface scanning image batches to obtain the matching processing coefficients corresponding to the five algorithms for each surface scanning image batch. After sorting these coefficients, the target algorithm matching heat map can be obtained. For example, for three surface scanning image batches, the target algorithm matching heat map obtained may be as follows: for surface scanning image batch 1, the matching processing coefficients corresponding to algorithm 1, algorithm 2, algorithm 3, algorithm 4, and algorithm 5 are 7.14, 3.5, 5.0, 8.0, and 2.0 respectively; for surface scanning image batch 2, the matching processing coefficients corresponding to the five algorithms are 4.0, 6.0, 3.0, 5.0, and 1.5 respectively; for surface scanning image batch 3, the matching processing coefficients corresponding to the five algorithms are 2.5, 3.0, 4.5, 6.0, and 1.0 respectively.

[0076] Finally, the linkage image algorithm index corresponding to each surface scan image batch is determined based on the target algorithm matching heat map. For each surface scan image batch, the corresponding row of matching processing coefficients is found in the target algorithm matching heat map.

[0077] Taking surface scan image batch 1 as an example, its corresponding matching processing coefficients are 7.14, 3.5, 5.0, 8.0, and 2.0. In this set of coefficients, 8.0 is the maximum value, and the algorithm corresponding to this maximum value is Algorithm 4. Therefore, the linkage image algorithm index corresponding to surface scan image batch 1 points to Algorithm 4.

[0078] Similarly, for surface scan image batch 2, the matching processing coefficients are 4.0, 6.0, 3.0, 5.0, and 1.5, among which 6.0 is the largest, and the corresponding algorithm is Algorithm 2. Therefore, the linkage image algorithm index corresponding to surface scan image batch 2 points to Algorithm 2.

[0079] For surface scanning image batch 3, the matching processing coefficients are 2.5, 3.0, 4.5, 6.0, and 1.0, among which 6.0 is the largest, and the corresponding algorithm is Algorithm 4. Therefore, the linkage image algorithm index corresponding to surface scanning image batch 3 points to Algorithm 4.

[0080] Through the above steps, the surface scanned images of enamel products in the target scanned image set were summarized more carefully based on the initial crack wavelet descriptor, and the surface scanned image batches and the linkage image algorithm index corresponding to each batch were determined. This method helps to improve the accuracy and pertinence of subsequent analysis and processing of surface cracks of enamel products, can better explore the characteristics of cracks, and provide a more reliable basis for quality inspection, protection and repair of enamel products.

[0081] In a replaceable embodiment, for each batch of surface scan images, the initial crack wavelet descriptor is subjected to multidimensional wavelet descriptor identification according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch to obtain the target edge morphology texture vector corresponding to the surface scan image batch, including: performing linear wavelet descriptor identification on the initial crack wavelet descriptor to obtain a crack linear wavelet descriptor; for each batch of surface scan images, the crack linear wavelet descriptor is input into the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch to obtain the corresponding first edge morphology texture vector, the first edge morphology texture vector is the target edge morphology texture vector, or the first edge morphology texture vector is a multidimensional reference vector used to obtain the target edge morphology texture vector.

[0082] In detail, for each batch of surface scanning images, a multidimensional wavelet descriptor is identified on the initial crack wavelet descriptor according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scanning image batch, so as to obtain the target edge morphology texture vector corresponding to the surface scanning image batch. The specific technical solution is as follows.

[0083] For example, the initial crack wavelet descriptor is a vector with a relatively high dimension, that is, a vector with a dimension of 1000, which contains various information about the surface cracks of enamel products, such as crack characteristics at different positions and scales.

[0084] When performing linear wavelet descriptor identification, an exemplary linear transformation method can be used. For example, a wavelet basis function with good time-frequency characteristics is selected, such as the Db4 wavelet basis in the Daubechies wavelet basis (this embodiment is only an example, and other suitable wavelet bases can also be selected). Each element in the initial crack wavelet descriptor is subjected to a linear transformation operation according to the rules of the Db4 wavelet basis.

[0085] For example, the element values ​​in the initial crack wavelet descriptor vector are (v_1, v_2, v_3, cdots, v_{1000}). After the linear transformation matrix (M) related to the Db4 wavelet basis (due to space reasons, the specific form of the matrix is ​​not expanded), a new vector is obtained. This new vector is the crack linear wavelet descriptor. For example, this crack linear wavelet descriptor is (w_1, w_2, w_3, cdots, w_{800}) (the dimension of this embodiment is changed to 800 just as an example, and the actual dimension is related to the wavelet basis and the transformation method). Compared with the initial crack wavelet descriptor, this crack linear wavelet descriptor retains the basic characteristic information of the crack and is presented in a form that is more suitable for subsequent processing. It highlights the linear relationship of crack characteristics at different scales.

[0086] Next, for example, a corresponding linkage image algorithm index has been determined for each surface scan image batch, and this index points to an exemplary image description processing algorithm. That is, there are three surface scan image batches, namely batch A, batch B, and batch C. For batch A, its linkage image algorithm index points to algorithm X; for batch B, its linkage image algorithm index points to algorithm Y; for batch C, its linkage image algorithm index points to algorithm Z.

[0087] These algorithms (algorithm X, algorithm Y, and algorithm Z) are designed for different types of crack features. For example, algorithm X may be an algorithm specifically designed for processing images with long linear crack features, algorithm Y may be better at processing cracks with more branching, and algorithm Z may focus on processing images with more complex crack edge textures.

[0088] Taking batch A as an example, the crack linear wavelet descriptor (w_1, w_2, w_3, cdots, w_{800}) corresponding to batch A is input into algorithm X.

[0089] Algorithm X may contain multiple processing steps. First, algorithm X may decompose the input crack linear wavelet descriptor, for example, decompose it into multiple sub-vectors, for example, decompose it into three sub-vectors (w_{1-300}, w_{301-600}, w_{601-800}). Then, feature extraction operations are performed on these sub-vectors respectively.

[0090] For the sub-vector (w_{1-300}), Algorithm X adopts an exemplary edge detection method. For example, this edge detection method is based on the gradient calculation method, which approximates the gradient by calculating the difference between adjacent elements in the sub-vector. For example, (w_2-w_1, w_3-w_2, cdots) is calculated to obtain a set of gradient values, and these gradient values ​​are combined in some way (such as summing, averaging, etc.) to obtain a value (e_1), which reflects the edge information of the crack under this part of the feature.

[0091] For the sub-vector (w_{301-600}), algorithm X uses a texture analysis method. For example, this method represents the texture feature by calculating the statistics of local features. For example, the variance of the elements within a certain range in the sub-vector is calculated to obtain a value (t_1), which reflects the texture information of the crack under this part of the feature.

[0092] For the subvector (w_{601-800}), algorithm X adopts a method of fusing edge and texture information. For example, the edge information (e_1) and texture information (t_1) obtained previously are combined according to certain weights, such as 0.6 and 0.4 respectively, to obtain a new value (f_1=0.6e_1+0.4t_1).

[0093] The values ​​(e_1, t_1, f_1) etc. obtained through the above steps are combined to form a new vector, which is the first edge morphology texture vector corresponding to batch A. For example, this vector is ([e_1, t_1, f_1, cdots]), which may contain 100 elements (this embodiment is only an example).

[0094] Similarly, for batch B and batch C, their corresponding crack linear wavelet descriptors are input into algorithm Y and algorithm Z respectively, and the first edge morphology texture vectors corresponding to batch B and batch C are obtained according to the processing logic within each algorithm.

[0095] In some cases, this first edge morphology texture vector is the required target edge morphology texture vector, which directly reflects the crack edge morphology and texture features corresponding to the surface scan image batch. In other cases, this first edge morphology texture vector may be a multi-dimensional reference vector used to obtain the target edge morphology texture vector, and further processing (such as feature fusion, adjustment, etc.) is required to obtain the final target edge morphology texture vector.

[0096] Through the above steps, for each batch of surface scanned images, the initial crack wavelet descriptor is processed according to the image description processing algorithm indicated by the corresponding linkage image algorithm index, and the corresponding first edge morphological texture vector is obtained, which provides an important feature basis for subsequent classification and evaluation of surface cracks of enamel products.

[0097] In addition, in the case where the first edge morphology texture vector is a multidimensional reference vector for obtaining the target edge morphology texture vector, for each batch of the surface scan images, the initial crack wavelet descriptor is subjected to multidimensional wavelet descriptor identification according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch to obtain the target edge morphology texture vector corresponding to the surface scan image batch, and further includes: performing linear wavelet descriptor identification on the first edge morphology texture vector corresponding to each of the surface scan image batches to obtain an edge morphology texture linear vector; for each batch of the surface scan images, the edge morphology texture linear vector is input into the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch to obtain the corresponding second edge morphology texture vector, the second edge morphology texture vector is the target edge morphology texture vector, or the second edge morphology texture vector is a multidimensional reference vector for obtaining the target edge morphology texture vector.

[0098] Further, in the case where the first edge morphology texture vector is a multidimensional reference vector for obtaining a target edge morphology texture vector, for each surface scanning image batch, a multidimensional wavelet descriptor recognition is performed on the initial crack wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scanning image batch to obtain a target edge morphology texture vector corresponding to the surface scanning image batch. There is also the following supplementary implementation method:

[0099] 1. Perform linear wavelet descriptor recognition on the first edge morphology texture vector corresponding to each batch of surface scanned images to obtain the edge morphology texture linear vector

[0100] For a certain batch of surface scanned images, the corresponding first edge morphology texture vector is a vector with a certain dimension, for example, a vector with a dimension of 200, whose element values ​​represent information related to the preliminary edge morphology and texture features of the cracks in the batch of images, such as (v_1, v_2, v_3, cdots, v_{200}).

[0101] When performing linear wavelet descriptor recognition, the embodiment of the present invention selects a suitable linear wavelet transform method, for example, using an exemplary wavelet basis function (such as Haar wavelet basis) for operation.

[0102] For each element in the first edge morphological texture vector, it is processed according to the transformation rule of the Haar wavelet basis. Specifically, the elements in the vector are grouped for transformation operations. Set every two elements as a group. For the first group (v_1) and (v_2), two new values ​​(w_1) and (w_2) are obtained after the Haar wavelet basis transformation (the transformation process in the embodiment of the present invention does not involve complex formula representation). The entire vector is processed in the same way to finally obtain a new vector, which is the edge morphological texture linear vector. Set this edge morphological texture linear vector to (w_1, w_2, w_3, cdots, w_{150}) (the dimension becoming 150 is just a possible example, depending on the wavelet basis and the transformation method). Compared with the first edge morphological texture vector, this edge morphological texture linear vector is presented in a form that is more suitable for further processing. It further explores the feature information in the first edge morphological texture vector, especially in terms of linear relationships at different scales.

[0103] 2. For each surface scan image batch, the edge morphology texture linear vector is input into the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scan image batch to obtain the corresponding second edge morphology texture vector

[0104] 1. Algorithm for linkage image algorithm index indication

[0105] The embodiment of the present invention has determined a linkage image algorithm index for each surface scanning image batch, and this index points to an exemplary image description processing algorithm. For example, for surface scanning image batch A, its linkage image algorithm index points to algorithm M; for surface scanning image batch B, its linkage image algorithm index points to algorithm N, etc. These algorithms are optimized for different crack feature types.

[0106] 2. Input the edge morphology texture linear vector into the algorithm to obtain the second edge morphology texture vector

[0107] Taking the surface scan image batch A as an example, its corresponding edge morphology texture linear vector (w_1, w_2, w_3, cdots, w_{150}) is input into the algorithm M.

[0108] Algorithm M has an exemplary processing logic inside. First, algorithm M may perform a segmentation operation on the input edge morphology texture linear vector, for example, segmenting it into several sub-vectors. Suppose the segmentation is into three sub-vectors (w_{1 - 50}, w_{51 - 100}, w_{101 - 150}).

[0109] For the sub-vector (w_{1 - 50}), Algorithm M uses a special edge refinement method. For example, by calculating the difference between the elements in the sub-vector, and screening and adjusting according to a certain threshold. For the adjacent elements (w_i) and (w_{i + 1}), the difference (d_i = w_{i + 1}-w_i) is calculated. If (d_i) is greater than a certain threshold (such as 0.5), (w_{i + 1}) is adjusted (for example, a specific value is subtracted) to obtain a new sub-vector element value. Through such operations, the edge feature information is further refined.

[0110] For the sub-vector (w_{51 - 100}), Algorithm M uses a texture feature enhancement method. For example, the representation of texture features is enhanced by performing weighted summation on the elements. Each element is assigned a weight (p_i) (the weight value can be determined based on experience or predefined rules), and then the weighted sum (s=sum_{i=51}^{100} p_i * w_i) is calculated to obtain a value representing the enhanced texture feature.

[0111] For the subvector (w_{101 - 150}), Algorithm M uses a new method to fuse edge and texture information. For example, the previously obtained refined edge features and enhanced texture features are fused at a certain ratio. The refined edge features are represented by (e_2), and the enhanced texture features are represented by (t_2). The fusion is performed at a ratio of 0.7 and 0.3 to obtain (f_2=0.7e_2 + 0.3t_2).

[0112] The various values ​​obtained through the above steps (such as the edge eigenvalue after refinement, the texture eigenvalue after enhancement, the eigenvalue after fusion, etc.) are combined to form a new vector, which is the second edge morphology texture vector corresponding to the surface scan image batch A. This vector is set to ([e_2, t_2, f_2, cdots]), which may have 80 elements (just an example in the embodiment of the present invention).

[0113] For other surface scan image batches, such as batch B, the corresponding edge morphology texture linear vectors are also input into algorithm N, and according to the processing logic of algorithm N, the second edge morphology texture vector corresponding to batch B is obtained.

[0114] In some cases, the second edge morphology texture vector is the target edge morphology texture vector required by the embodiment of the present invention, which fully and accurately reflects the crack edge morphology and texture characteristics corresponding to the surface scanning image batch. In other cases, the second edge morphology texture vector may still be a multi-dimensional reference vector for obtaining the target edge morphology texture vector, and further processing (such as performing similar transformation, fusion, etc. again) may be required to finally obtain the target edge morphology texture vector.

[0115] Through the above steps, when the first edge morphology texture vector is a multidimensional reference vector, each surface scan image batch is further processed to gradually and deeply explore the edge morphology and texture characteristics of the cracks to obtain a more accurate target edge morphology texture vector, which will help to perform more accurate classification and evaluation of surface cracks on enamel products in the future.

[0116] In an optional embodiment, the method is implemented according to an enamel crack sorting network, the enamel crack sorting network includes a processing algorithm matching branch, an initial wavelet descriptor recognition branch, a multidimensional wavelet descriptor recognition branch and a sorting label decision branch, the multidimensional wavelet descriptor recognition branch includes several of the image description processing algorithms, and the enamel crack sorting network is debugged through the following steps: obtaining a historical scan image set consisting of several historical enamel product surface scan images; inputting the historical scan image set into the initial wavelet descriptor recognition branch to obtain an initial historical crack wavelet descriptor; inputting the initial historical crack wavelet descriptor into the processing algorithm matching branch to obtain several historical surface scan image batches and linkage image algorithm indexes corresponding to each of the historical surface scan image batches; for each of the historical surface scan image batches, according to the linkage image algorithm corresponding to the multidimensional wavelet descriptor recognition branch, The image description processing algorithm corresponding to the index performs multidimensional wavelet descriptor recognition on the initial historical crack wavelet descriptor to obtain the historical edge morphology texture vector corresponding to the historical surface scanning image batch; based on the historical edge morphology texture vectors corresponding to each of the historical surface scanning image batches, a historical crack sorting decision wavelet descriptor is obtained; the historical crack sorting decision wavelet descriptor is input into the sorting label decision branch to obtain the enamel crack sorting training label; a network training offset is determined, the network training offset includes a first debugging error for indicating the difference between the enamel crack sorting training label and prior training information, the prior training information indicates the enamel crack sorting category of the historical scanning image set; based on the network training offset, the neural network variables of the processing algorithm matching branch, the initial wavelet descriptor identification branch, the multidimensional wavelet descriptor identification branch and the sorting label decision branch are optimized.

[0117] In practical application, this method can be implemented through an enamel crack sorting network, which includes a processing algorithm matching branch, an initial wavelet descriptor recognition branch, a multidimensional wavelet descriptor recognition branch, and a sorting label decision branch, wherein the multidimensional wavelet descriptor recognition branch covers several image description processing algorithms. The following is a detailed description of the debugging steps of the enamel crack sorting network.

[0118] 1. Get a collection of historical scanned images

[0119] First, we need to obtain a set of historical scanned images consisting of several scanned images of the surface of historical enamel products. These historical scanned images are collected by specific equipment (such as high-resolution scanning equipment), and contain detailed information on the surface of enamel products of different historical periods and types. They are the basic data source for debugging the enamel crack sorting network.

[0120] 2. Obtain the initial historical crack wavelet descriptor in the initial wavelet descriptor identification branch

[0121] The historical scan image set is input into the initial wavelet descriptor identification branch. The initial wavelet descriptor identification branch uses wavelet analysis technology to process the input historical scan image. For example, an algorithm based on discrete wavelet transform (DWT) can be used, which transforms the pixel values ​​in the historical scan image to extract the initial historical crack wavelet descriptor that can characterize the crack characteristics. Specifically, the DWT algorithm decomposes the image into sub-bands of different frequencies, and finally forms the initial historical crack wavelet descriptor by combining and processing the coefficients of the low-frequency and high-frequency sub-bands. This descriptor is a vector containing multiple elements, each of which represents the characteristic information related to the crack at different positions and frequencies in the image.

[0122] 3. Get batches and indexes in the algorithm matching branch

[0123] The initial historical crack wavelet descriptor is input into the processing algorithm matching branch. The processing algorithm matching branch uses a clustering algorithm (such as the K-means algorithm) to process the initial historical crack wavelet descriptor. The K-means algorithm uses the distance between data points as a similarity metric to cluster the initial historical crack wavelet descriptors into several classes, each of which corresponds to a batch of historical surface scan images. For example, if the number of cluster centers is set to k (k is a predefined integer determined according to actual conditions), the K-means algorithm will iteratively assign the initial historical crack wavelet descriptor to the nearest cluster center, and after multiple iterations, a stable batch of k historical surface scan images will be formed.

[0124] For each batch of historical surface scanned images, the processing algorithm matching branch determines the corresponding linkage image algorithm index based on the characteristics of the batch. This determination process can be based on a predefined algorithm feature matching rule. For example, the feature distribution of the initial historical crack wavelet descriptor of each batch is compared with the applicable features of several pre-set image description processing algorithms, the best matching image description processing algorithm is selected, and the index corresponding to the algorithm is assigned to this batch.

[0125] 4. Obtain the historical edge morphology texture vector in the multidimensional wavelet descriptor recognition branch

[0126] For each batch of historical surface scanned images, the initial historical crack wavelet descriptor is identified by a multidimensional wavelet descriptor according to the image description processing algorithm corresponding to the corresponding linkage image algorithm index in the multidimensional wavelet descriptor identification branch. For example, the image description processing algorithm corresponding to a linkage image algorithm index may be an algorithm based on multi-scale wavelet transform (MSWT).

[0127] The MSWT algorithm will perform a more detailed decomposition and analysis of the initial historical crack wavelet descriptor. It performs wavelet transform on the initial historical crack wavelet descriptor at multiple scales, extracts feature information such as the edge and texture of the crack at different scales, such as by calculating the statistical characteristics of the wavelet coefficients at different scales (such as mean, variance, etc.), and then combines these feature information to finally obtain the historical edge morphology texture vector corresponding to the batch of historical surface scan images. This vector is a vector containing rich crack edge morphology and texture feature information, and its elements are values ​​related to crack features calculated by a specific algorithm.

[0128] 5. Obtain the historical crack sorting decision wavelet descriptor

[0129] Based on the historical edge morphology texture vectors corresponding to each batch of historical surface scanning images, the historical crack sorting decision wavelet descriptor is obtained. The specific operation can be to splice each historical edge morphology texture vector in a certain order. For example, if there are n batches of historical surface scanning images, and the historical edge morphology texture vector corresponding to each batch is (V_i) ((i=1, 2, cdots, n)), then the historical crack sorting decision wavelet descriptor (D) can be formed by splicing these vectors in sequence, that is, (D= (V_1, V_2, cdots, V_n)), and the historical crack sorting decision wavelet descriptor obtained in this way contains the crack feature information of all historical surface scanning image batches.

[0130] 6. Obtain enamel crack sorting training labels in the sorting label decision branch

[0131] The historical crack sorting decision wavelet descriptor is input into the sorting label decision branch. The sorting label decision branch uses a classification algorithm (such as the support vector machine (SVM) algorithm) to process the input historical crack sorting decision wavelet descriptor. The SVM algorithm separates different categories of data (different types of enamel crack features) by finding a hyperplane. Based on the feature information in the historical crack sorting decision wavelet descriptor, the SVM algorithm determines the enamel crack sorting training label corresponding to each historical scan image, for example, enamel cracks are divided into different categories such as linear cracks and reticular cracks, and each category corresponds to a specific label.

[0132] 7. Determine network training offset

[0133] The network training offset includes a first debugging error for indicating the difference between the enamel crack sorting training label and the prior training information. The prior training information indicates the enamel crack sorting category of the historical scan image set, which is a known correct classification result. The first debugging error can be calculated using an error calculation method, such as a cross-entropy error function. Assuming the enamel crack sorting training label is (y_{pred}), and the correct label in the prior training information is (y_{true}), then the cross-entropy error function (E=- sum_{i}y_{true}^ilog(y_{pred}^i)), (i) in the embodiment of the present invention represents different samples (i.e., historical scan images). The error (E) obtained by calculation is the first debugging error, which reflects the degree of difference between the enamel crack sorting training label output by the current network and the correct label.

[0134] 8. Optimize Neural Network Variables of Network Branches

[0135] According to the network training offset, the neural network variables of the processing algorithm matching branch, the initial wavelet descriptor identification branch, the multidimensional wavelet descriptor identification branch and the sorting label decision branch are optimized. For example, for the initial wavelet descriptor identification branch, if the DWT-based algorithm is used, some parameters in the wavelet transform (such as the selection of wavelet basis functions, etc.) may be adjusted according to the first debugging error. For the processing algorithm matching branch, if the K-means algorithm is used, the initialization method of the cluster center or the stop condition of the iteration may be adjusted. For the multidimensional wavelet descriptor identification branch, if the MSWT algorithm is used, the number of layers of the multiscale decomposition or the method of feature extraction at different scales may be adjusted. For the sorting label decision branch, if the SVM algorithm is used, the parameters of the kernel function or the weight of the support vector may be adjusted. By optimizing these neural network variables, the enamel crack sorting network can more accurately sort the cracks on the surface of enamel products.

[0136] In some other optional embodiments, the processing algorithm matching branch includes a category weight activation module and an algorithm matching analysis module, and the inputting of the initial historical crack wavelet descriptor into the processing algorithm matching branch to obtain a number of historical surface scanning image batches and a linkage image algorithm index corresponding to each of the historical surface scanning image batches includes: inputting the initial historical crack wavelet descriptor into the category weight activation module to obtain a historical crack attention embedding vector and a preliminary screening batch category weight corresponding to each of the historical enamel product surface scanning images determined based on the historical crack attention embedding vector; determining the number of historical surface scanning image batches based on the preliminary screening batch category weight corresponding to each of the historical enamel product surface scanning images; determining the corresponding historical image batch cross description vector in the historical crack attention embedding vector for each of the historical surface scanning image batches; inputting the historical image batch cross description vector into the algorithm matching analysis module to obtain a historical algorithm matching heat map; and determining the linkage image algorithm index corresponding to the historical surface scanning image batch based on the historical algorithm matching heat map.

[0137] Given that the processing algorithm matching branch includes a category weight activation module and an algorithm matching parsing module, the following is a specific implementation method of inputting the initial historical crack wavelet descriptor into the processing algorithm matching branch to obtain several historical surface scanning image batches and the linkage image algorithm index corresponding to each batch.

[0138] 1. Operations in the Class Weight Activation Module

[0139] The initial historical crack wavelet descriptor is input into the category weight activation module. The category weight activation module uses a specific attention mechanism algorithm to process the input initial historical crack wavelet descriptor. This attention mechanism algorithm is designed to highlight the key features related to batch classification in the initial historical crack wavelet descriptor. For example, an attention mechanism based on a convolutional neural network (CNN) is used. The convolution layer of the CNN extracts features from the initial historical crack wavelet descriptor, and then converts these features into a historical crack attention embedding vector through a fully connected layer. This historical crack attention embedding vector is a vector containing feature information filtered and enhanced by the attention mechanism, and its element values ​​represent the quantitative representation of important features related to surface cracks of enamel products.

[0140] At the same time, the weight of the initial screening batch category corresponding to each historical enamel product surface scan image is determined based on the historical crack attention embedding vector. The process of determining the weight of the initial screening batch category can adopt an algorithm based on distance measurement. For example, for each historical enamel product surface scan image corresponding to the historical crack attention embedding vector part, calculate its distance from several predefined category center vectors (these category center vectors are determined in advance based on historical data or prior knowledge, representing different potential batch categories). The distance calculation in the embodiment of the present invention can adopt a variant of the Euclidean distance algorithm, and the distance is expressed by calculating the sum of the absolute values ​​of the differences between the vectors. According to this distance, the weight of the initial screening batch category corresponding to each historical enamel product surface scan image is determined by a mapping function (for example, a linear mapping function with a higher weight the closer the distance). These weights reflect the degree of possibility that each historical enamel product surface scan image belongs to each potential batch category.

[0141] 2. Determine the historical surface scan image batch

[0142] According to the weight of the initial screening batch category corresponding to each historical enamel product surface scan image, several batches of historical surface scan images are determined. Specifically, a threshold segmentation algorithm can be used. For example, a weight threshold is set. If the weight of a certain initial screening batch category corresponding to a certain historical enamel product surface scan image exceeds this threshold, the image is divided into the batch of historical surface scan images corresponding to this initial screening batch category. By performing such an operation on all historical enamel product surface scan images, several batches of historical surface scan images can be obtained. The images in these batches have similarities in the initial screening batch category weights, which means that they have certain commonalities in the features related to batch classification.

[0143] 3. Determine the cross description vector of historical image batches for historical surface scanned image batches

[0144] For each batch of historical surface scanned images, the corresponding cross description vector of the historical image batch is determined in the historical crack attention embedding vector. For each batch of historical surface scanned images, a feature selection algorithm can be used. For example, a feature selection algorithm based on information gain. First, the information gain value of each element in the historical crack attention embedding vector for the current batch of historical surface scanned images is calculated. This information gain value can be calculated using a statistical method by comparing the degree of reduction in uncertainty in batch classification when a certain element is included and when the element is not included. Then, several elements with large information gain values ​​are selected, and these elements are extracted from the historical crack attention embedding vector to form a cross description vector of the historical image batch. This vector is a subset of the historical crack attention embedding vector, which contains the most valuable feature information for the classification of the current batch of historical surface scanned images.

[0145] 4. Operations in the Algorithm Matching and Parsing Module

[0146] The cross description vector of the historical image batch is input into the algorithm matching and parsing module. The algorithm matching and parsing module uses a predefined algorithm matching algorithm to process the input cross description vector of the historical image batch. For example, an algorithm matching algorithm based on cosine similarity is used. For each cross description vector of the historical image batch, the cosine similarity between the feature template vector corresponding to several predefined image description processing algorithms (these feature template vectors are determined in advance according to the characteristics of the algorithm and the requirements for different types of crack processing) is calculated.

[0147] Through the above calculations, the historical algorithm matching heat map is obtained. The historical algorithm matching heat map is a two-dimensional representation, in which the rows represent different historical surface scan image batches and the columns represent different image description processing algorithms. The value of each element represents the cosine similarity between the cross description vector of the corresponding historical image batch and the feature template vector of the image description processing algorithm. This similarity is the matching processing coefficient in the historical algorithm matching heat map, which reflects the matching degree of each image description processing algorithm for each historical surface scan image batch.

[0148] 5. Determine the linkage image algorithm index

[0149] According to the historical algorithm matching heat map, determine the linkage image algorithm index corresponding to the historical surface scan image batch. For each historical surface scan image batch, find its corresponding row in the historical algorithm matching heat map, and then find the element with the largest matching processing coefficient in this row. The index of the image description processing algorithm corresponding to this maximum matching processing coefficient is the linkage image algorithm index corresponding to the historical surface scan image batch. This index will be used to select the appropriate image description processing algorithm in the multidimensional wavelet descriptor identification branch to further process the initial historical crack wavelet descriptor, so as to obtain a more accurate historical edge morphology texture vector, thereby providing support for the accurate operation of the entire enamel crack sorting network.

[0150] In another optional example, the network training offset also includes a second debugging error, and the method also includes: obtaining a preset statistical value, the preset statistical value indicating the authenticated number of batches of historical surface scanning images; performing preliminary screening processing on the preliminary screening batch category weights corresponding to each of the historical enamel product surface scanning images to obtain a category weight screening error; and determining the second debugging error based on the difference between the category weight screening error and the preset statistical value.

[0151] Based on the above embodiment, the second debugging error in the network training offset is determined as follows: First, a preset statistical value is obtained. This preset statistical value is used to indicate the certified number of historical surface scanning image batches. For example, this preset statistical value may be an integer, such as 100, indicating the number of historical surface scanning image batches determined by a certain certification standard in the entire historical data set. This certification standard may be determined based on previous experiments, prior knowledge, or other reliable basis.

[0152] Then, the initial screening batch category weights corresponding to the surface scan images of each historical enamel product are initially screened to obtain the category weight screening error. The initial screening process may involve a series of analytical operations on the initial screening batch category weights. For example, in the initial screening process, there may be a predefined weight range or weight mode. For each surface scan image of a historical enamel product, its initial screening batch category weight is set to a numerical value, and these values ​​are distributed between 0 and 1. If a certain rule is followed, such as treating weights less than 0.3 or greater than 0.7 as outliers (the actual rule is determined according to specific technical requirements), and then calculating the difference between the actual number of outliers and the number of outliers expected according to the ideal model, this difference is the category weight screening error.

[0153] Finally, the second debugging error is determined based on the difference between the category weight screening error and the preset statistical value. Specifically, if the category weight screening error is (e) and the preset statistical value is (n), then the second debugging error (E) can be determined by a certain calculation method. For example, a simple difference calculation can be used, that is, (E = e - n), or a more complex calculation based on a proportional relationship can be used, such as (E=frac{e}{n}) (this example is only to illustrate the calculation relationship and does not mean that only these two calculation methods can be used in reality). This calculation method reflects the degree of deviation of the category weight screening error from the preset statistical value, thereby determining the second debugging error.

[0154] In this way, the way in which the second debugging error is determined in this embodiment helps to more comprehensively evaluate the performance of the enamel crack sorting network. By considering the relationship between the number of certified batches of historical surface scanning images and the screening error of the initial screening batch category weights, the accuracy of the network in batch classification and weight processing can be more accurately measured. This helps to discover possible problems in the network, such as whether the processing of category weights during the initial screening process is reasonable, and then provides a more targeted basis for optimizing related modules in the network (such as processing algorithm matching branches, etc.), thereby improving the accuracy and reliability of the entire enamel crack sorting network in sorting surface cracks of enamel products.

[0155] In addition, the network training offset also includes a third debugging error, and the historical algorithm matching heat map includes the matching hit score corresponding to the linkage image algorithm index corresponding to the historical surface scan image batch. Based on this, the method also includes: determining a corresponding influence weight for each linkage image algorithm index, the influence weight indicating the ratio of the total number of historical enamel product surface scan images corresponding to the linkage image algorithm index to the total number of all historical enamel product surface scan images in the historical scan image set; determining the timeliness evaluation corresponding to the linkage image algorithm index based on the weighted calculation result of the influence weight and the matching hit score corresponding to the linkage image algorithm index; determining the third debugging error based on the timeliness evaluation corresponding to each linkage image algorithm index.

[0156] In an embodiment of an embodiment of the present invention, for the third debugging error in the network training offset, it is determined as follows: First, for each linkage image algorithm index, the corresponding influence weight is determined. The influence weight in the embodiment of the present invention is determined based on the ratio of the total number of historical enamel product surface scan images corresponding to the linkage image algorithm index to the total number of all historical enamel product surface scan images in the historical scan image set. For example, assuming that there are a total of 1,000 historical enamel product surface scan images in the historical scan image set, and there are 200 historical enamel product surface scan images corresponding to a certain linkage image algorithm index, then the influence weight corresponding to this linkage image algorithm index is 200 / 1000=0.2. This influence weight reflects the relative importance of the linkage image algorithm index in the entire historical scan image set.

[0157] Next, the timeliness evaluation corresponding to the linkage image algorithm index is determined based on the weighted calculation result of the influence weight and the matching hit score corresponding to the linkage image algorithm index. Assuming that the matching hit score of a certain linkage image algorithm index is 0.8 (this matching hit score is obtained from the historical algorithm matching heat map, which reflects the matching degree between the algorithm index and the corresponding historical surface scan image batch), and its corresponding influence weight is 0.2 as calculated above, then their weighted calculation result is 0.2×0.8=0.16, and this 0.16 is the timeliness evaluation corresponding to the linkage image algorithm index. This timeliness evaluation comprehensively considers the matching degree of the algorithm index and its importance in the entire historical scan image collection.

[0158] Finally, the third debugging error is determined based on the timeliness evaluation corresponding to each linkage image algorithm index. For example, if there are 5 linkage image algorithm indexes, and their corresponding timeliness evaluations are 0.16, 0.12, 0.18, 0.09, and 0.15, respectively, then the third debugging error can be determined by performing some comprehensive calculation on these timeliness evaluations. One possible way is to calculate the difference between the average value of these timeliness evaluations and a preset standard value (this standard value is determined based on historical experience or theoretical analysis, such as 0.15). Assuming the calculated average value is 0.14, the third debugging error is 0.14-0.15=-0.01.

[0159] It can be seen that the way this embodiment determines the third debugging error helps to comprehensively evaluate the effectiveness of the linkage image algorithm index in the enamel crack sorting network. By considering the impact weight and matching hit score to determine the timeliness evaluation, and then deriving the third debugging error, the performance of each linkage image algorithm index in the entire sorting network can be accurately measured. This helps to discover whether the algorithm index has timeliness problems in different batches of images, and provides a basis for optimizing the algorithm selection and matching mechanism in the network, thereby improving the accuracy and stability of the entire network in sorting cracks on the surface of enamel products.

[0160] In some independent embodiments, after processing the crack sorting decision wavelet descriptor to obtain the enamel crack sorting label, the method further includes: performing abnormal monitoring on the enamel product production line matched by the target scanning image set based on the enamel crack sorting label to obtain abnormal monitoring results; and performing inspection and maintenance on the enamel product production line based on the abnormal monitoring results.

[0161] In detail, based on the enamel crack sorting label, the enamel product production line matched by the target scanning image set is monitored for abnormalities to obtain abnormal monitoring results, including: performing state trend identification on the first sensor monitoring data stream of the enamel product production line to obtain a first state trend feature stream corresponding to the first sensor monitoring data stream; performing outlier feature screening based on the first state trend feature stream to obtain a screened second state trend feature stream; performing production line abnormality discrimination on the second state trend feature stream to obtain a first production line abnormality discrimination result corresponding to the second state trend feature stream.

[0162] In addition, after obtaining the first production line abnormality judgment result corresponding to the second state trend characteristic flow, the method also includes: performing outlier feature screening based on at least part of the sensor monitoring data in the first sensor monitoring data flow to obtain the screened second production line abnormality judgment result; obtaining a key equipment abnormality judgment result based on the first production line abnormality judgment result and the second production line abnormality judgment result.

[0163] In detail, the state trend identification is performed on the first sensor monitoring data stream of the enamel product production line to obtain the first state trend feature stream corresponding to the first sensor monitoring data stream, including: one-hot encoding is performed on the first sensor monitoring data stream of the enamel product production line to obtain the first one-hot encoded vector stream corresponding to the first sensor monitoring data stream; attention processing is performed on the first one-hot encoded vector stream to obtain the first state trend feature stream corresponding to the first sensor monitoring data stream.

[0164] Furthermore, the performing of production line abnormality identification on the second state trend characteristic stream to obtain a first production line abnormality identification result corresponding to the second state trend characteristic stream includes: performing production line abnormality identification on the second state trend characteristic stream to obtain a second one-hot encoded vector stream corresponding to the second state trend characteristic stream; performing decoding processing based on the spatiotemporal domain on the second one-hot encoded vector stream to obtain a first production line abnormality identification result corresponding to the second state trend characteristic stream.

[0165] Furthermore, the production line abnormality judgment is performed on the second state trend feature stream to obtain a second one-hot encoded vector stream corresponding to the second state trend feature stream, including: using a fully connected layer to perform cyclic full-connection processing on the second state trend feature stream to obtain a second one-hot encoded vector stream corresponding to the second state trend feature stream.

[0166] In some more detailed technical solutions, the outlier features are screened according to the first state trend feature stream to obtain a screened second state trend feature stream, including: inputting the first state trend feature stream into a deep residual model, and outputting a second state trend feature stream corresponding to the first state trend feature stream via the deep residual model.

[0167] It can be understood that after obtaining the enamel crack sorting label, it involves performing abnormal monitoring on the enamel product production line matched by the target scan image set according to the label and performing maintenance based on the results. The specific implementation method is as follows:

[0168] 1. Abnormal monitoring of enamel product production line based on enamel crack sorting label to obtain abnormal monitoring results

[0169] 1. Perform state trend recognition on the first sensor monitoring data stream of the enamel product production line to obtain the first state trend characteristic stream

[0170] First, the first sensor monitoring data stream of the enamel product production line is one-hot encoded to obtain the first one-hot encoded vector stream corresponding to the first sensor monitoring data stream. For example, assuming that the first sensor monitoring data stream contains multiple different types of sensor data, such as temperature, pressure, humidity and other data, the one-hot encoding will convert each different type of data value into a specific vector representation. If the temperature has 5 different value ranges, then the one-hot encoding will convert each temperature value into a 5-dimensional vector, in which only one dimension has a value of 1, and the rest are 0, indicating the range to which the temperature value belongs. In this way, after one-hot encoding each data type in the entire first sensor monitoring data stream, the first one-hot encoded vector stream is obtained.

[0171] Then, the first one-hot encoded vector stream is processed with attention to obtain the first state trend feature stream corresponding to the first sensor monitoring data stream. The attention processing mechanism will weight the vectors in the first one-hot encoded vector stream to highlight those feature vectors that have an important impact on the state trend. For example, in the production process, the temperature data may have a greater impact on the state trend of the production line. In this way, the attention mechanism will assign a higher weight to the vector corresponding to the temperature. In this way, the first one-hot encoded vector stream is converted into the first state trend feature stream, which can better reflect the state trend of the enamel product production line.

[0172] 2. Screen outlier features based on the first state trend feature flow to obtain the screened second state trend feature flow

[0173] The first state trend feature stream is input into the deep residual model, and the second state trend feature stream corresponding to the first state trend feature stream is output through the deep residual model. The deep residual model has multiple residual blocks, and each residual block can learn the complex relationship in the data. For example, each residual block in the deep residual model can analyze the features in the first state trend feature stream, remove those outlier features that deviate greatly from the normal state trend, and thus obtain the second state trend feature stream. This deep residual model can effectively reduce the impact of noise and outliers in the data on subsequent judgments.

[0174] 3. Perform production line abnormality discrimination on the second state trend characteristic flow to obtain the first production line abnormality discrimination result corresponding to the second state trend characteristic flow

[0175] First, a fully connected layer is used to perform cyclic full connection processing on the second state trend feature stream to obtain the second unique hot encoding vector stream corresponding to the second state trend feature stream. The neurons in the fully connected layer will be connected to all neurons in the previous layer, and the cyclic full connection processing can further integrate and transform the features in the second state trend feature stream. For example, the fully connected layer will perform weighted summation and other operations on each feature in the second state trend feature stream according to the pre-set weights, and convert it into a second unique hot encoding vector stream, which can represent the state characteristics of the production line in a new form.

[0176] Then, the second one-hot encoded vector stream is decoded based on the time and space domain to obtain the first production line abnormality judgment result corresponding to the second state trend feature stream. The decoding process based on the time and space domain takes into account the relationship between the data in time and space. For example, if the production line data in a certain area fluctuates abnormally over a period of time, the time and space domain decoding process will comprehensively consider the data characteristics within this time and space range to determine whether there is a production line abnormality, and finally obtain the first production line abnormality judgment result, which may indicate that the production line is normal or there is some type of abnormality (such as equipment failure, process problems, etc.).

[0177] 4. Perform outlier feature screening based on at least part of the sensor monitoring data in the first sensor monitoring data stream to obtain the screened second production line abnormality identification result

[0178] Select at least part of the sensor monitoring data from the first sensor monitoring data stream, for example, select those data that have a greater impact on the stability of the production line operation, such as the temperature and pressure data of key equipment. Statistical methods can be used to screen outlier features for these selected data, such as calculating the mean and standard deviation of the data, and treating those data that deviate from the mean by a certain multiple of the standard deviation as outliers. In this way, the abnormality discrimination result of the second production line is obtained, which is also a judgment on whether there is an abnormality in the production line.

[0179] 5. Based on the abnormality identification results of the first production line and the second production line, the abnormality identification results of the key equipment are obtained.

[0180] The abnormality identification results of the first production line and the abnormality identification results of the second production line are combined to determine the abnormality identification results of the key equipment. For example, if the abnormality identification results of the first production line indicate that the production line has an overall process abnormality, and the abnormality identification results of the second production line show that there are outliers in the temperature data of the key equipment, then it can be judged that the key equipment may have a fault, thereby obtaining the abnormality identification results of the key equipment. This result will provide an important basis for the subsequent inspection and maintenance of the enamel product production line.

[0181] 2. Inspection and maintenance of enamel product production lines based on abnormal monitoring results

[0182] Once the critical equipment abnormality identification results are obtained, if the results show that there is an equipment abnormality, a corresponding inspection and maintenance plan can be formulated according to the type and severity of the abnormality. For example, if the critical equipment abnormality identification results show that the temperature of a critical equipment is too high, there may be a problem with the cooling system, then the inspection and maintenance plan may include inspection, cleaning or replacement of parts of the cooling system. If it is a process abnormality, it may be necessary to adjust the production process, such as adjusting production parameters, checking the quality of raw materials, etc.

[0183] The beneficial effect of this embodiment is that by performing multi-step processing on the sensor monitoring data stream of the enamel product production line, including state trend identification, outlier feature screening and abnormality discrimination, the operating status of the production line can be fully and accurately monitored. Combined with the analysis based on enamel crack sorting labels, abnormal conditions in the production line can be discovered in a timely manner, especially abnormalities of key equipment. This helps to take maintenance measures in advance, reduce product quality problems caused by production line failures, improve the efficiency and quality of enamel product production, reduce production costs, and ensure the stability and reliability of the entire enamel product production process.

[0184] Based on the same inventive concept, an embodiment of the present invention provides an enamel crack sorting device, comprising:

[0185] An image recognition module is used to obtain a target scan image set consisting of a plurality of scan images of the surface of enamel products; perform wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor;

[0186] An image summarization module is used to summarize the surface scan images of the enamel products in the target scan image set according to the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, each of the surface scan image batches including a plurality of surface scan images of the enamel products, and the linkage image algorithm index is an index corresponding to an image description processing algorithm for implementing multidimensional wavelet descriptor recognition;

[0187] an algorithm matching module, for performing multidimensional wavelet descriptor recognition on the initial crack wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scanning image batch for each of the surface scanning image batches, to obtain a target edge morphology texture vector corresponding to the surface scanning image batch;

[0188] The crack sorting module is used to splice the target edge morphology texture vectors corresponding to each batch of the surface scanned images to obtain the crack sorting decision wavelet descriptor corresponding to the target scanned image set; and process the crack sorting decision wavelet descriptor to obtain the enamel crack sorting label.

[0189] Further, Figure 2 FIG. 2 is a schematic diagram of a computer device 200 provided in an embodiment of the present invention. Figure 2 The computer device 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method in the embodiment of the present invention.

[0190] Alternatively, if Figure 2 As shown, the computer device 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention.

[0191] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0192] Alternatively, if Figure 2 As shown, the computer device 200 may further include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices, specifically, to send information or data to other devices, or to receive information or data sent by other devices.

[0193] Optionally, the computer device 200 may implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present invention, which will not be described in detail here for the sake of brevity.

[0194] It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capability.

[0195] It is understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0196] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0197] It should be noted that, in this article, the term "comprises", "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0198] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned implementation methods can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the embodiment of the present invention.

[0199] The embodiments of the embodiments of the present invention are described above in conjunction with the accompanying drawings, but the embodiments of the present invention are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present invention and the scope of protection of the embodiments of the present invention, all of which are within the protection of the embodiments of the present invention.

Claims

1. A method for sorting enamel cracks, characterized in that: The method is implemented by a computer device, and the method comprises: Acquire a target scan image set consisting of a plurality of scan images of the surface of enamel products; perform wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor; According to the initial crack wavelet descriptor, the surface scan images of the enamel products in the target scan image set are summarized to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to the surface scan image batches, each of the surface scan image batches includes a plurality of the surface scan images of the enamel products, and the linkage image algorithm index is an index corresponding to the image description processing algorithm for implementing multidimensional wavelet descriptor recognition; For each of the surface scanning image batches, performing multidimensional wavelet descriptor recognition on the initial crack wavelet descriptor according to the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scanning image batch, and obtaining the target edge morphology texture vector corresponding to the surface scanning image batch; splicing the target edge morphology texture vectors corresponding to each of the surface scan image batches to obtain a crack sorting decision wavelet descriptor corresponding to the target scan image set; processing the crack sorting decision wavelet descriptor to obtain an enamel crack sorting label; The method is implemented according to an enamel crack sorting network, which includes a processing algorithm matching branch, an initial wavelet descriptor identification branch, a multidimensional wavelet descriptor identification branch and a sorting label decision branch. The multidimensional wavelet descriptor identification branch includes several image description processing algorithms. The enamel crack sorting network is debugged through the following steps: Acquire a historical scan image set consisting of a plurality of historical enamel product surface scan images; Inputting the historical scanned image set into the initial wavelet descriptor identification branch to obtain an initial historical crack wavelet descriptor; Inputting the initial historical crack wavelet descriptor into the processing algorithm matching branch to obtain a plurality of historical surface scanning image batches and linkage image algorithm indexes corresponding to each of the historical surface scanning image batches; For each batch of historical surface scan images, multidimensional wavelet descriptor recognition is performed on the initial historical crack wavelet descriptor according to the image description processing algorithm corresponding to the corresponding linkage image algorithm index in the multidimensional wavelet descriptor recognition branch to obtain the historical edge morphology texture vector corresponding to the batch of historical surface scan images; Based on the historical edge morphology texture vectors corresponding to each batch of historical surface scanned images, a historical crack sorting decision wavelet descriptor is obtained; Inputting the historical crack sorting decision wavelet descriptor into the sorting label decision branch to obtain enamel crack sorting training labels; determining a network training offset, the network training offset comprising a first debugging error indicating a difference between the enamel crack sorting training label and prior training information indicating enamel crack sorting categories of the historical scan image set; According to the network training offset, the neural network variables of the processing algorithm matching branch, the initial wavelet descriptor identification branch, the multidimensional wavelet descriptor identification branch and the sorting label decision branch are optimized.

2. The method according to claim 1, characterized in that The method of summarizing the surface scanned images of the enamel products in the target scanned image set based on the initial crack wavelet descriptor to obtain a plurality of surface scanned image batches and linkage image algorithm indexes corresponding to each of the surface scanned image batches includes: Performing attention feature embedding on the initial crack wavelet descriptor to obtain a crack attention embedding vector; Identifying the initial screening batch category weight corresponding to each of the enamel product surface scan images according to the crack attention embedding vector; By performing a preliminary screening process on the category weights of each of the preliminary screening batches, at least one candidate enamel product surface scan image is determined, where the candidate enamel product surface scan image is an enamel product surface scan image whose corresponding category weight meets a set condition; Based on the surface scan images of each candidate enamel product, the plurality of surface scan image batches are obtained.

3. The method according to claim 2, characterized in that The method further comprises: summarizing the surface scan images of the enamel products in the target scan image set according to the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to the surface scan image batches; For each of the surface scan image batches, determining a corresponding image batch cross description vector in the crack attention embedding vector; Determining a target algorithm matching heat map according to the image batch cross description vector, wherein the target algorithm matching heat map indicates a matching processing coefficient corresponding to each image description processing algorithm; According to the target algorithm matching heat map, the linkage image algorithm index corresponding to the surface scanning image batch is determined.

4. The method according to any one of claims 1 to 3, characterized in that For each of the surface scan image batches, the initial crack wavelet descriptor is identified with a multidimensional wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scan image batch, to obtain a target edge morphology texture vector corresponding to the surface scan image batch, including: Performing linear wavelet descriptor identification on the initial crack wavelet descriptor to obtain a crack linear wavelet descriptor; For each batch of surface scanning images, the crack linear wavelet descriptor is input into the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scanning image batch to obtain the corresponding first edge morphology texture vector, where the first edge morphology texture vector is the target edge morphology texture vector, or the first edge morphology texture vector is a multidimensional reference vector used to obtain the target edge morphology texture vector.

5. The method according to claim 4, characterized in that In the case where the first edge morphology texture vector is a multidimensional reference vector for obtaining the target edge morphology texture vector, the step of performing multidimensional wavelet descriptor identification on the initial crack wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scanning image batch for each of the surface scanning image batches to obtain the target edge morphology texture vector corresponding to the surface scanning image batch further includes: Performing linear wavelet descriptor recognition on the first edge morphology texture vectors corresponding to each of the surface scan image batches to obtain edge morphology texture linear vectors; For each batch of surface scanned images, the edge morphological texture linear vector is input into the image description processing algorithm indicated by the linkage image algorithm index corresponding to the surface scanned image batch to obtain the corresponding second edge morphological texture vector, where the second edge morphological texture vector is the target edge morphological texture vector, or the second edge morphological texture vector is a multidimensional reference vector used to obtain the target edge morphological texture vector.

6. The method according to claim 1, characterized in that The processing algorithm matching branch includes a category weight activation module and an algorithm matching parsing module. The initial historical crack wavelet descriptor is input into the processing algorithm matching branch to obtain a plurality of historical surface scanning image batches and linkage image algorithm indexes corresponding to each of the historical surface scanning image batches, including: Inputting the initial historical crack wavelet descriptor into the category weight activation module to obtain a historical crack attention embedding vector and a primary screening batch category weight corresponding to each of the historical enamel product surface scan images determined according to the historical crack attention embedding vector; Determining the batches of historical surface scan images according to the initial screening batch category weight corresponding to each of the historical enamel product surface scan images; For each of the historical surface scan image batches, determining a corresponding historical image batch cross description vector in the historical crack attention embedding vector; Inputting the cross description vector of the historical image batch into the algorithm matching analysis module to obtain a historical algorithm matching heat map; Determining the linkage image algorithm index corresponding to the batch of historical surface scan images according to the historical algorithm matching heat map; The network training offset also includes a second debugging error, and the method further includes: Obtaining a preset statistical value, the preset statistical value indicating the number of authentications of a batch of historical surface scan images; Performing a preliminary screening process on the category weights of the preliminary screening batches corresponding to the surface scanned images of the historical enamel products to obtain a category weight screening error; Determining the second debugging error according to the difference between the category weight screening error and the preset statistical value; The network training offset further includes a third debugging error, the historical algorithm matching heat map includes a matching hit score corresponding to the linkage image algorithm index corresponding to the historical surface scanning image batch, and the method further includes: For each linkage image algorithm index, determining a corresponding influence weight, the influence weight indicating a ratio of the total number of historical enamel product surface scan images corresponding to the linkage image algorithm index to the total number of all historical enamel product surface scan images in the historical scan image set; Determining the timeliness evaluation corresponding to the linkage image algorithm index according to a weighted calculation result of the influence weight and the matching hit score corresponding to the linkage image algorithm index; The third debugging error is determined according to the timeliness evaluation corresponding to each of the linkage image algorithm indexes.

7. An enamel crack sorting device, characterized in that: include: An image recognition module is used to obtain a target scan image set consisting of a plurality of scan images of the surface of enamel products; perform wavelet descriptor recognition on the target scan image set to obtain an initial crack wavelet descriptor; An image summarization module is used to summarize the surface scan images of the enamel products in the target scan image set according to the initial crack wavelet descriptor to obtain a plurality of surface scan image batches and linkage image algorithm indexes corresponding to each of the surface scan image batches, each of the surface scan image batches including a plurality of surface scan images of the enamel products, and the linkage image algorithm index is an index corresponding to an image description processing algorithm for implementing multidimensional wavelet descriptor recognition; an algorithm matching module, for performing multidimensional wavelet descriptor recognition on the initial crack wavelet descriptor according to an image description processing algorithm indicated by a linkage image algorithm index corresponding to the surface scanning image batch for each of the surface scanning image batches, to obtain a target edge morphology texture vector corresponding to the surface scanning image batch; A crack sorting module is used to splice the target edge morphology texture vectors corresponding to each batch of the surface scanned images to obtain a crack sorting decision wavelet descriptor corresponding to the target scanned image set; and process the crack sorting decision wavelet descriptor to obtain an enamel crack sorting label; The image recognition module, the image summarization module, the algorithm matching module and the crack sorting module are implemented according to the enamel crack sorting network. The enamel crack sorting network includes a processing algorithm matching branch, an initial wavelet descriptor identification branch, a multidimensional wavelet descriptor identification branch and a sorting label decision branch. The multidimensional wavelet descriptor identification branch includes several image description processing algorithms. The enamel crack sorting network is debugged through the following steps: Acquire a historical scan image set consisting of a plurality of historical enamel product surface scan images; Inputting the historical scanned image set into the initial wavelet descriptor identification branch to obtain an initial historical crack wavelet descriptor; Inputting the initial historical crack wavelet descriptor into the processing algorithm matching branch to obtain a plurality of historical surface scanning image batches and linkage image algorithm indexes corresponding to each of the historical surface scanning image batches; For each batch of historical surface scan images, multidimensional wavelet descriptor recognition is performed on the initial historical crack wavelet descriptor according to the image description processing algorithm corresponding to the corresponding linkage image algorithm index in the multidimensional wavelet descriptor recognition branch to obtain the historical edge morphology texture vector corresponding to the batch of historical surface scan images; Based on the historical edge morphology texture vectors corresponding to each batch of historical surface scanned images, a historical crack sorting decision wavelet descriptor is obtained; Inputting the historical crack sorting decision wavelet descriptor into the sorting label decision branch to obtain enamel crack sorting training labels; determining a network training offset, the network training offset comprising a first debugging error indicating a difference between the enamel crack sorting training label and prior training information indicating enamel crack sorting categories of the historical scan image set; According to the network training offset, the neural network variables of the processing algorithm matching branch, the initial wavelet descriptor identification branch, the multidimensional wavelet descriptor identification branch and the sorting label decision branch are optimized.

8. A computer device, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 6.

9. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images