Intelligent manufacturing method and system for stainless steel cabinets

Abstract

The present application relates to the field of intelligent manufacturing detection, and specifically discloses an intelligent manufacturing method and system for stainless steel cabinets, which adopts a convolutional neural network model based on deep learning technology to extract implicit correlation features from multiple perspective images of assembled stainless steel cabinets collected from multiple perspectives from both local and global aspects, and uses a layered deep homography alignment fusion method to perform homography alignment of layered scene depth streams based on vector differential expression during feature fusion, so as to improve the classification effect of the fused features, and thus can accurately detect the assembly quality of the assembled stainless steel cabinets.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing detection, and more specifically, to an intelligent manufacturing method and system for stainless steel cabinets. Background Art

[0002] Today, people's demands for living environments and quality are constantly increasing. From living rooms to bedrooms to kitchens, the transformation of every aspect reflects people's demand for a better quality of life. For the entire kitchen, stainless steel cabinets, as the main component, combined with modern products, can better reflect people's taste for life. Due to the special nature of the kitchen, which cannot be easily replaced, the quality of the cabinets becomes particularly important. Therefore, the quality inspection of stainless steel cabinets after manufacturing and assembly is also extremely important.

[0003] However, after the existing stainless steel cabinets are assembled, most of them need to be inspected by personal inspection equipment, which is not only time-consuming and costly, but also leads to low inspection accuracy. Therefore, an optimized assembly quality inspection method for stainless steel cabinets is desired. Summary of the Invention

[0004] In order to solve the above technical problems, the present application is proposed. The embodiments of the present application provide an intelligent manufacturing method and system for stainless steel cabinets, which adopt a convolutional neural network model based on deep learning technology to extract implicit correlation features from multiple perspective images of assembled stainless steel cabinets collected from multiple perspectives in both local and global aspects, and use a layered deep homography alignment fusion method to perform homography alignment of layered scene depth streams based on vector differential expression during feature fusion, so as to improve the classification effect of the fused features, and thus can accurately detect the assembly quality of the assembled stainless steel cabinets.

[0005] According to one aspect of the present application, a method for intelligently manufacturing a stainless steel cabinet is provided, comprising:

[0006] The camera collects multiple perspective images of the assembled stainless steel cabinet from multiple perspectives;

[0007] Passing each of the multiple perspective images through a first convolutional neural network as a feature extractor to obtain multiple local perspective feature maps;

[0008] Arranging the multiple local view feature maps into a three-dimensional input tensor and obtaining an inter-view correlation feature map through a second convolutional neural network using a three-dimensional convolution kernel;

[0009] Performing panoramic stitching on the multiple local view feature maps to obtain a panoramic feature map;

[0010] Passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map;

[0011] Expanding the inter-view correlation feature map and the global correlation feature map into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector respectively;

[0012] fusing the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector; and

[0013] The classification feature vector is passed through a classifier to obtain a classification result, and the classification result is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements.

[0014] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, each of the multiple perspective images is passed through a first convolutional neural network as a feature extractor to obtain multiple local perspective feature maps, including: each layer of the first convolutional neural network performs the following on the input data in the forward pass of the layer: convolution processing on the input data to obtain a convolution feature map; pooling processing on the convolution feature map to obtain a pooling feature map; and nonlinear activation on the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the multiple local perspective feature maps, and the input of the first layer of the first convolutional neural network is each of the multiple perspective images.

[0015] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, after arranging the multiple local perspective feature maps into a three-dimensional input tensor, a second convolutional neural network using a three-dimensional convolution kernel is used to obtain an inter-perspective correlation feature map, including: the second convolutional neural network using the three-dimensional convolution kernel performs the following on the input data in the forward pass of the layer: performing three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; performing pooling processing on the convolution feature map to obtain a pooling feature map; and, performing nonlinear activation on the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the second convolutional neural network is the inter-perspective correlation feature map, and the input of the first layer of the second convolutional neural network is the three-dimensional input tensor.

[0016] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the multiple local perspective feature maps are panoramically stitched to obtain a panoramic feature map, including: stitching the multiple local perspective feature maps along the height dimension or the width dimension to obtain the panoramic feature map.

[0017] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the panoramic feature map is passed through a non-local neural network to obtain a global correlation feature map, including: inputting the panoramic feature map into the first point convolution layer, the second point convolution layer and the third point convolution layer of the non-local neural network respectively to obtain a first feature map, a second feature map and a third feature map; calculating the position-weighted sum of the first feature map and the second feature map to obtain an intermediate fusion feature map; inputting the intermediate fusion feature map into a Softmax function to normalize the feature values of each position in the intermediate fusion feature map to obtain a normalized intermediate fusion feature map; calculating the position-weighted sum of the normalized intermediate fusion feature map and the third feature map to obtain a re-fusion feature map; embedding the re-fusion feature map through a Gaussian similarity function to calculate the similarity between the feature values of each position in the re-fusion feature map to obtain a global perception feature map; passing the global perception feature map through the fourth point convolution layer of the non-local neural network to obtain a channel-adjusted global perception feature map; and calculating the position-weighted sum of the channel-adjusted global perception feature map and the high-dimensional correlation local feature map to obtain the global correlation feature map.

[0018] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the inter-view correlation feature map and the global correlation feature map are respectively expanded into vectors to obtain the inter-view correlation feature vector and the global correlation feature vector, including: dividing each feature matrix in the inter-view correlation feature map and each feature matrix in the global correlation feature map with row vectors as division units to obtain multiple first row vectors corresponding to the inter-view correlation feature map and multiple second row vectors corresponding to the global correlation feature map; splicing the multiple first row vectors to obtain the inter-view correlation feature vector; and splicing the multiple second row vectors to obtain the global correlation feature vector.

[0019] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the inter-view correlation feature vector and the global correlation feature vector are fused to obtain a classification feature vector, including: calculating the position difference of the inter-view correlation feature vector and the global correlation feature vector to obtain a differential feature vector; calculating the position sum of the inter-view correlation feature vector and the global correlation feature vector to obtain a sum feature vector; calculating the full-scene homography association matrix between the inter-view correlation feature vector and the global correlation feature vector; performing a logarithmic operation on the sum feature vector to obtain a logarithmic sum feature vector, wherein the performing a logarithmic operation on the sum feature vector The row logarithm operation represents calculating the logarithmic function value of the eigenvalue of each position in the summed eigenvector; calculating the norm of the differential eigenvector as the hierarchical depth characteristic value of the differential eigenvector; calculating the Frobenius norm of the full-scene homography association matrix as the depth perception value of the full-scene homography association matrix; weighting the logarithmic summed eigenvector by position using the hierarchical depth characteristic value as a weighting coefficient to obtain a weighted eigenvector; and, using the depth perception value as a bias, calculating the sum of the depth perception value and the eigenvalue of each position in the weighted eigenvector to obtain the classified eigenvector.

[0020] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the full-scene homography association matrix between the inter-view correlation feature vector and the global correlation feature vector is calculated, including: calculating the product between the transposed vector of the inter-view correlation feature vector and the global correlation feature vector to obtain the full-scene homography association matrix.

[0021] In the above-mentioned intelligent manufacturing method of stainless steel cabinets, the classification feature vector is passed through a classifier to obtain a classification result, including: using the classifier to process the classification feature vector using the following formula to obtain the classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|X}, where W1 to W n is the weight matrix, B1 to B n is the bias vector, and X is the classification feature vector.

[0022] According to another aspect of the present application, there is provided an intelligent manufacturing system for a stainless steel cabinet, comprising:

[0023] An image data acquisition unit, configured to capture images of the assembled stainless steel cabinet from multiple viewing angles using a camera;

[0024] a local feature extraction unit, configured to pass each of the plurality of perspective images through a first convolutional neural network serving as a feature extractor to obtain a plurality of local perspective feature maps;

[0025] A three-dimensional convolution unit is configured to arrange the multiple local view feature maps into a three-dimensional input tensor and obtain an inter-view correlation feature map by using a second convolutional neural network with a three-dimensional convolution kernel;

[0026] a panoramic stitching unit, configured to perform panoramic stitching on the multiple local view feature maps to obtain a panoramic feature map;

[0027] A global feature extraction unit, configured to pass the panoramic feature map through a non-local neural network to obtain a global correlation feature map;

[0028] a dimensionality reduction unit, configured to expand the inter-view correlation feature map and the global correlation feature map into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector;

[0029] a feature fusion unit, configured to fuse the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector; and

[0030] The classification unit is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets the predetermined requirements.

[0031] In the above-mentioned intelligent manufacturing system of stainless steel cabinets, the local feature extraction unit is further used to: each layer of the first convolutional neural network performs the following on the input data in the forward pass of the layer: convolution processing on the input data to obtain a convolution feature map; pooling processing on the convolution feature map to obtain a pooling feature map; and nonlinear activation on the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the multiple local perspective feature maps, and the input of the first layer of the first convolutional neural network is each perspective image in the multiple perspective images.

[0032] In the above-mentioned intelligent manufacturing system of stainless steel cabinets, the three-dimensional convolution unit is further used to: the second convolutional neural network using the three-dimensional convolution kernel performs the following on the input data in the forward pass of the layer: performing three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; performing pooling processing on the convolution feature map to obtain a pooled feature map; and, performing nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the second convolutional neural network is the inter-view correlation feature map, and the input of the first layer of the second convolutional neural network is the three-dimensional input tensor.

[0033] In the above-mentioned intelligent manufacturing system for stainless steel cabinets, the panoramic stitching unit is further used to stitch the multiple local perspective feature maps along the height dimension or the width dimension to obtain the panoramic feature map.

[0034] In the above-mentioned intelligent manufacturing system of stainless steel cabinets, the global feature extraction unit is further used to: input the panoramic feature map into the first point convolution layer, the second point convolution layer and the third point convolution layer of the non-local neural network respectively to obtain a first feature map, a second feature map and a third feature map; calculate the position-weighted sum of the first feature map and the second feature map to obtain an intermediate fusion feature map; input the intermediate fusion feature map into the Softmax function to normalize the feature values of each position in the intermediate fusion feature map to obtain a normalized intermediate fusion feature map; calculate the position-weighted sum of the normalized intermediate fusion feature map and the third feature map to obtain a re-fusion feature map; embed the re-fusion feature map through a Gaussian similarity function to calculate the similarity between the feature values of each position in the re-fusion feature map to obtain a global perception feature map; pass the global perception feature map through the fourth point convolution layer of the non-local neural network to obtain a channel-adjusted global perception feature map; and calculate the position-weighted sum of the channel-adjusted global perception feature map and the high-dimensional correlation local feature map to obtain the global correlation feature map.

[0035] In the above-mentioned intelligent manufacturing system of stainless steel cabinets, the dimensionality reduction unit is further used to: divide each feature matrix in the inter-view correlation feature map and each feature matrix in the global correlation feature map with row vectors as division units to obtain multiple first row vectors corresponding to the inter-view correlation feature map and multiple second row vectors corresponding to the global correlation feature map; splice the multiple first row vectors to obtain the inter-view correlation feature vector; and splice the multiple second row vectors to obtain the global correlation feature vector.

[0036] In the above-mentioned intelligent manufacturing system of stainless steel cabinets, the feature fusion unit includes: a differential subunit, which is used to calculate the positional difference of the inter-perspective correlation feature vector and the global correlation feature vector to obtain a differential feature vector; a summation subunit, which is used to calculate the positional summation of the inter-perspective correlation feature vector and the global correlation feature vector to obtain a summation feature vector; an association matrix subunit, which is used to calculate the full-scene homography association matrix between the inter-perspective correlation feature vector and the global correlation feature vector; a logarithmic calculation subunit, which is used to perform a logarithmic operation on the summation feature vector to obtain a logarithmic summation feature vector, wherein the logarithmic operation on the summation feature vector represents the calculation of the summation feature vector. The present invention relates to a method for obtaining a classification feature vector comprising: a first derivative of ...

[0037] In the above-mentioned intelligent manufacturing system for stainless steel cabinets, the correlation matrix subunit is further used to: calculate the product between the transposed vector of the inter-view correlation feature vector and the global correlation feature vector to obtain the full-scene homography correlation matrix.

[0038] In the above-mentioned intelligent manufacturing system for stainless steel cabinets, the classification unit is further used to: use the classifier to process the classification feature vector according to the following formula to obtain the classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|X}, where W1 to W n is the weight matrix, B1 to B n is the bias vector, and X is the classification feature vector.

[0039] According to another aspect of the present application, a computer-readable medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the processor executes the intelligent manufacturing method for stainless steel cabinets as described above.

[0040] Compared with the existing technology, the intelligent manufacturing method and system of stainless steel cabinets provided in the present application adopt a convolutional neural network model based on deep learning technology to extract implicit correlation features from multiple perspective images of assembled stainless steel cabinets collected from multiple perspectives from both local and global aspects, and use a layered deep homography alignment fusion method to perform homography alignment of layered scene depth streams based on vector differential expression during feature fusion, so as to improve the classification effect of the fused features, and thus can accurately detect the assembly quality of the assembled stainless steel cabinets. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0042] Figure 1 This is a diagram showing an application scenario of the intelligent manufacturing method for a stainless steel cabinet according to an embodiment of the present application;

[0043] Figure 2 Flowchart of a method for intelligent manufacturing of a stainless steel cabinet according to an embodiment of the present application;

[0044] Figure 3 Schematic diagram of the system architecture of the intelligent manufacturing method of stainless steel cabinets according to an embodiment of the present application;

[0045] Figure 4 A flowchart of fusing the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector in the intelligent manufacturing method for a stainless steel cabinet according to an embodiment of the present application;

[0046] Figure 5 is a block diagram of an intelligent manufacturing system for stainless steel cabinets according to an embodiment of the present application;

[0047] Figure 6 This is a block diagram of a feature fusion unit in the intelligent manufacturing system of a stainless steel cabinet according to an embodiment of the present application. DETAILED DESCRIPTION

[0048] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described herein.

[0049] Scenario Overview

[0050] As mentioned earlier, people's demands for living environments and quality are constantly increasing. From living rooms to bedrooms to kitchens, the transformation of every aspect reflects people's demand for a better quality of life. For the entire kitchen, stainless steel cabinets, as the main component, combined with modern products, can better reflect people's taste for life. Due to the special nature of the kitchen, which cannot be easily replaced, the quality of the cabinets becomes particularly important, and thus the quality inspection of stainless steel cabinets after manufacturing and assembly becomes extremely important.

[0051] However, after the existing stainless steel cabinets are assembled, most of them need to be inspected by personal inspection equipment, which is not only time-consuming and costly, but also leads to low inspection accuracy. Therefore, an optimized assembly quality inspection method for stainless steel cabinets is desired.

[0052] Accordingly, the inventors of this application, taking into account the development of deep neural networks, attempted to perform quality inspection of cabinet assembly only through images without the need for various testing equipment to perform tests such as structural strength verification, thereby reducing quality inspection costs.

[0053] Based on this, in the technical solution of this application, a camera is first used to capture multiple perspective images of an assembled stainless steel cabinet from multiple viewpoints. These multiple perspective images are then processed using a convolutional neural network model, which excels in image feature extraction, to extract local implicit feature distribution information from the multiple perspective images, thereby obtaining multiple local perspective feature maps.

[0054] It should be understood that, considering that in the multiple perspective images of the assembled stainless steel cabinet, each perspective image has mutually correlated feature information, therefore, in order to better extract the feature information about the assembly quality of the stainless steel cabinet from the multiple perspective images, the multiple local perspective feature maps are further arranged into a three-dimensional input tensor and then a second convolutional neural network using a three-dimensional convolution kernel is used to obtain a perspective correlation feature map. Here, the second convolutional neural network can use the three-dimensional convolution kernel to deeply mine the spatial feature information of the assembled stainless steel cabinet.

[0055] Furthermore, considering that convolution is a typical local operation, and for each pixel point in the multiple perspective images of the assembled stainless steel cabinet, the pixels do not exist in isolation, and the correlation between pixels produces a foreground target. Therefore, in the technical solution of the present application, in order to extract the correlation between a certain pixel point and all the remaining pixels in the multiple perspective images of the assembled stainless steel cabinet, a non-local neural network is used to further perform image feature extraction. That is, the multiple local perspective feature maps are panoramically spliced to obtain a panoramic feature map, and the panoramic feature map is passed through a non-local neural network to obtain a global correlation feature map. In particular, here, the non-local neural network captures long-range dependency information by calculating the similarity of all pixels in the image, and then models contextual features, so that the network pays attention to the overall content of the image, thereby improving the feature extraction capability of the backbone network in classification and detection tasks.

[0056] Furthermore, when fusing the inter-view correlation feature map and the global correlation feature map, considering that the inter-view correlation feature map itself performs feature extraction through the cascaded first and second convolutional neural networks, and the global correlation feature map further increases the layer depth through the convolution operation of the non-local neural network on the basis of the inter-view correlation feature map, therefore, in the technical solution of the present application, when fusing the inter-view correlation feature map and the global correlation feature map, layered depth homography alignment fusion is performed.

[0057] Specifically, the inter-view correlation feature map and the global correlation feature map are first expanded into an inter-view correlation feature vector V1 and a global correlation feature vector V2, and then the hierarchical deep homography alignment fusion of the feature vectors is performed, which is expressed as:

[0058]

[0059] Where V1 represents the inter-view correlation feature vector, V2 represents the global correlation feature vector, V3 represents the classification feature vector, ||·||1 represents the norm of the vector, and ||·|| F represents the Frobenius norm of the matrix, and denote positional subtraction and addition, respectively, and ⊙ denotes positional dot multiplication.

[0060] Here, the layered depth homography alignment fusion performs homography alignment of the layered scene depth stream based on vector differential expression according to the layered depth characteristics based on feature fusion represented by the vector, and uses the depth perception of the full scene homography correlation matrix between the vectors as the superimposed bias. In this way, the layered depth homography alignment of the features is effectively performed on the basis of the distribution misalignment caused by different layered depth features between the feature distributions of the inter-perspective correlation feature map and the global correlation feature map, thereby improving the classification effect of the fused classification feature vector V3, so as to improve the accuracy of assembly quality inspection of the assembled stainless steel cabinets.

[0061] Based on this, the present application proposes an intelligent manufacturing method for stainless steel cabinets, which includes: collecting multiple perspective images of the assembled stainless steel cabinet from multiple perspectives through a camera; passing each of the multiple perspective images through a first convolutional neural network serving as a feature extractor to obtain multiple local perspective feature maps; arranging the multiple local perspective feature maps into a three-dimensional input tensor and then passing it through a second convolutional neural network using a three-dimensional convolution kernel to obtain an inter-perspective correlation feature map; performing panoramic stitching on the multiple local perspective feature maps to obtain a panoramic feature map; passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map; expanding the inter-perspective correlation feature map and the global correlation feature map into vectors respectively to obtain an inter-perspective correlation feature vector and a global correlation feature vector; fusing the inter-perspective correlation feature vector and the global correlation feature vector to obtain a classification feature vector; and passing the classification feature vector through a classifier to obtain a classification result, which is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements.

[0062] Figure 1 The figure shows an application scenario diagram of the intelligent manufacturing method of a stainless steel cabinet according to an embodiment of the present application. Figure 1 As shown, in this application scenario, first, through a camera (for example, Figure 1 C) captures the assembled stainless steel cabinet from multiple perspectives (e.g., Figure 1 Then, the obtained multiple perspective images of the stainless steel cabinet are input into a server deployed with an intelligent manufacturing algorithm for the stainless steel cabinet (for example, Figure 1 S) as shown in , wherein the server is capable of processing multiple perspective images of the stainless steel cabinet using an intelligent manufacturing algorithm for the stainless steel cabinet to generate a classification result indicating whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements.

[0063] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0064] Exemplary Methods

[0065] Figure 2 The figure shows a flow chart of the intelligent manufacturing method of stainless steel cabinets. Figure 2 As shown, the intelligent manufacturing method of a stainless steel cabinet according to an embodiment of the present application includes: S110, collecting multiple perspective images of the assembled stainless steel cabinet from multiple perspectives through a camera; S120, passing each perspective image of the multiple perspective images through a first convolutional neural network as a feature extractor to obtain multiple local perspective feature maps; S130, arranging the multiple local perspective feature maps into a three-dimensional input tensor and then passing it through a second convolutional neural network using a three-dimensional convolution kernel to obtain an inter-perspective correlation feature map; S140, performing panoramic stitching on the multiple local perspective feature maps to obtain a panoramic feature map; S150, passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map; S160, respectively expanding the inter-perspective correlation feature map and the global correlation feature map into vectors to obtain an inter-perspective correlation feature vector and a global correlation feature vector; S170, fusing the inter-perspective correlation feature vector and the global correlation feature vector to obtain a classification feature vector; and, S180, passing the classification feature vector through a classifier to obtain a classification result, which is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements.

[0066] Figure 3 The figure shows a schematic diagram of the architecture of the intelligent manufacturing method of the stainless steel cabinet according to the embodiment of the present application. Figure 3 As shown, in the network architecture of the intelligent manufacturing method of the stainless steel cabinet, first, the multiple perspective images (for example, Figure 3 Each perspective image in IN0) is respectively processed by a first convolutional neural network (e.g., Figure 3 CNN1 as shown in the figure) to obtain multiple local view feature maps (for example, Figure 3 Then, the plurality of local view feature maps are arranged into a three-dimensional input tensor (e.g., as Figure 3 ) is then processed by a second convolutional neural network (e.g., Figure 3 CNN2 as shown in the figure) to obtain the inter-view correlation feature map (for example, Figure 3 Then, the plurality of local view feature maps are panorama stitched together to obtain a panorama feature map (for example, Figure 3 Then, the panoramic feature map is passed through a non-local neural network (e.g., Figure 3 CNN3 as shown in ) to obtain a global correlation feature map (e.g., Figure 3 Then, the inter-view correlation feature map and the global correlation feature map are respectively expanded into vectors to obtain an inter-view correlation feature vector (for example, Figure 3 VF1) and the global correlation feature vector (e.g., Figure 3 Then, the inter-view correlation feature vector and the global correlation feature vector are fused to obtain a classification feature vector (for example, Figure 3 and, finally, passing the classification feature vector through a classifier (e.g., Figure 3 The circle S) shown in FIG is used to obtain a classification result, which is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets the predetermined requirements.

[0067] In step S110 and step S120, a plurality of perspective images of the assembled stainless steel cabinet are collected from a plurality of perspectives by a camera; and each of the plurality of perspective images is passed through a first convolutional neural network as a feature extractor to obtain a plurality of local perspective feature maps. As mentioned above, in the technical solution of the present application, it is expected that the quality inspection of the cabinet assembly can be performed only by images without the need for various inspection equipment to perform inspections such as structural strength verification, so as to reduce the cost of quality inspection. Specifically, in the technical solution of the present application, first, a plurality of perspective images of the assembled stainless steel cabinet are collected from a plurality of perspectives by a camera. Then, the images of the plurality of perspectives are subjected to feature extraction processing in a convolutional neural network model with excellent performance in image feature extraction, so as to extract the local implicit feature distribution information in the plurality of perspective images, thereby obtaining a plurality of local perspective feature maps.

[0068] Specifically, in an embodiment of the present application, the process of passing each of the multiple perspective images through a first convolutional neural network as a feature extractor to obtain a plurality of local perspective feature maps includes: each layer of the first convolutional neural network performs the following on the input data in the forward pass of the layer: convolution processing on the input data to obtain a convolution feature map; pooling processing on the convolution feature map to obtain a pooling feature map; and nonlinear activation on the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the multiple local perspective feature maps, and the input of the first layer of the first convolutional neural network is each of the multiple perspective images.

[0069] In step S130, the multiple local perspective feature maps are arranged into a three-dimensional input tensor and then a second convolutional neural network with a three-dimensional convolution kernel is used to obtain a perspective correlation feature map. It should be understood that, considering that in the multiple perspective images of the assembled stainless steel cabinet, each perspective image has mutually correlated feature information, therefore, in order to better extract the feature information about the assembly quality of the stainless steel cabinet from the multiple perspective images, in the technical solution of the present application, the multiple local perspective feature maps are further arranged into a three-dimensional input tensor and then a second convolutional neural network with a three-dimensional convolution kernel is used to obtain a perspective correlation feature map. Here, the second convolutional neural network can use a three-dimensional convolution kernel to deeply mine the spatial feature information of the assembled stainless steel cabinet.

[0070] Specifically, in an embodiment of the present application, the process of arranging the multiple local view feature maps into a three-dimensional input tensor and then obtaining an inter-view correlation feature map by using a second convolutional neural network with a three-dimensional convolution kernel includes: the second convolutional neural network using the three-dimensional convolution kernel performs the following on the input data in the forward pass of the layer: performing three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; performing pooling processing on the convolution feature map to obtain a pooled feature map; and performing nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the second convolutional neural network is the inter-view correlation feature map, and the input of the first layer of the second convolutional neural network is the three-dimensional input tensor.

[0071] In steps S140 and S150, the multiple local perspective feature maps are panoramically stitched to obtain a panoramic feature map, and the panoramic feature map is passed through a non-local neural network to obtain a global correlation feature map. It should be understood that, considering that convolution is a typical local operation, the pixels in the multiple perspective images of the assembled stainless steel cabinet do not exist in isolation, and the correlation between pixels produces a foreground target. Therefore, in the technical solution of the present application, in order to extract the correlation between a certain pixel and all the remaining pixels in the multiple perspective images of the assembled stainless steel cabinet, a non-local neural network is used to further extract the features of the image. That is, the multiple local perspective feature maps are panoramically stitched to obtain a panoramic feature map. Accordingly, in a specific example, the multiple local perspective feature maps can be stitched along the height dimension or the width dimension to obtain the panoramic feature map. Then, the panoramic feature map is passed through a non-local neural network to obtain a global correlation feature map. In particular, here, the non-local neural network captures long-range dependency information by calculating the similarity of all pixels in the image, and then models contextual features, allowing the network to focus on the overall content of the image, thereby improving the backbone network's feature extraction capabilities in classification and detection tasks.

[0072] Specifically, in an embodiment of the present application, the process of passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map includes: first, inputting the panoramic feature map into the first point convolution layer, the second point convolution layer and the third point convolution layer of the non-local neural network respectively to obtain a first feature map, a second feature map and a third feature map; then, calculating the position-weighted sum of the first feature map and the second feature map to obtain an intermediate fusion feature map; then, inputting the intermediate fusion feature map into a Softmax function to normalize the feature values of each position in the intermediate fusion feature map to obtain a normalized intermediate feature map. Then, the normalized intermediate fusion feature map is fused; then, the position-weighted sum of the normalized intermediate fusion feature map and the third feature map is calculated to obtain a re-fused feature map; then, the re-fused feature map is embedded in a Gaussian similarity function to calculate the similarity between the feature values of each position in the re-fused feature map to obtain a global perception feature map; then, the global perception feature map is passed through the fourth point convolution layer of the non-local neural network to obtain a channel-adjusted global perception feature map; and finally, the position-weighted sum of the channel-adjusted global perception feature map and the high-dimensional correlation local feature map is calculated to obtain the global correlation feature map.

[0073] In step S160 and step S170, the inter-view correlation feature map and the global correlation feature map are respectively expanded into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector, and the inter-view correlation feature vector and the global correlation feature vector are fused to obtain a classification feature vector. It should be understood that, further, when fusing the inter-view correlation feature map and the global correlation feature map, considering that the inter-view correlation feature map itself is subjected to feature extraction through the cascaded first and second convolutional neural networks, and the global correlation feature map further increases the layer depth through the convolution operation of the non-local neural network on the basis of the inter-view correlation feature map, therefore, in the technical solution of the present application, when fusing the inter-view correlation feature map and the global correlation feature map, hierarchical deep homography alignment fusion is performed. Specifically, the inter-view correlation feature map and the global correlation feature map are first expanded into an inter-view correlation feature vector V1 and a global correlation feature vector V2, and then the hierarchical deep homography alignment fusion of the feature vectors is performed.

[0074] It should be understandable that the layered depth homography alignment fusion performs homography alignment of the layered scene depth stream based on vector differential expression according to the layered depth characteristics based on feature fusion represented by the vector, and uses the depth perception of the full scene homography correlation matrix between the vectors as the superimposed bias. In this way, the layered depth homography alignment of the features is effectively performed on the basis of the distribution misalignment between the feature distributions of the inter-perspective correlation feature map and the global correlation feature map caused by different layered depth features, thereby improving the classification effect of the fused classification feature vector V3, so as to improve the accuracy of assembly quality inspection of the assembled stainless steel cabinets.

[0075] Specifically, in an embodiment of the present application, the process of expanding the inter-view correlation feature map and the global correlation feature map into vectors respectively to obtain the inter-view correlation feature vector and the global correlation feature vector includes: dividing each feature matrix in the inter-view correlation feature map and each feature matrix in the global correlation feature map with row vectors as division units to obtain multiple first row vectors corresponding to the inter-view correlation feature map and multiple second row vectors corresponding to the global correlation feature map; splicing the multiple first row vectors to obtain the inter-view correlation feature vector; and splicing the multiple second row vectors to obtain the global correlation feature vector.

[0076] Specifically, in an embodiment of the present application, the process of fusing the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector includes: first, calculating the positional difference between the inter-view correlation feature vector and the global correlation feature vector to obtain a differential feature vector. Next, calculating the positional sum of the inter-view correlation feature vector and the global correlation feature vector to obtain a sum feature vector. Then, calculating the full-scene homography correlation matrix between the inter-view correlation feature vector and the global correlation feature vector. Correspondingly, in a specific example, the product between the transposed vector of the inter-view correlation feature vector and the global correlation feature vector is calculated to obtain the full-scene homography correlation matrix. Next, performing a logarithmic operation on the sum feature vector to obtain a logarithmic sum feature vector, wherein performing a logarithmic operation on the sum feature vector means calculating the logarithmic function value of the eigenvalues at each position in the sum feature vector. Then, calculating the norm of the differential feature vector as the hierarchical depth feature value of the differential feature vector. Next, the Frobenius norm of the full-scene homography correlation matrix is calculated as the depth perception value of the full-scene homography correlation matrix. Then, the logarithmic sum eigenvector is weighted by position using the layered depth characteristic value as a weighting coefficient to obtain a weighted eigenvector. Finally, the depth perception value is used as a bias to calculate the sum of the depth perception value and the eigenvalues of each position in the weighted eigenvector to obtain the classification eigenvector. That is, in a specific example, the formula for fusing the inter-view correlation eigenvector and the global correlation eigenvector is expressed as:

[0077]

[0078] Where V1 represents the inter-view correlation feature vector, V2 represents the global correlation feature vector, V3 represents the classification feature vector, ||·||1 represents the norm of the vector, and ||·|| F represents the Frobenius norm of the matrix, and denote positional subtraction and addition, respectively, and ⊙ denotes positional dot multiplication.

[0079] Figure 4 The figure shows a flow chart of fusing the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector in the intelligent manufacturing method of a stainless steel cabinet according to an embodiment of the present application. Figure 4As shown, in an embodiment of the present application, the inter-view correlation feature vector and the global correlation feature vector are fused to obtain a classification feature vector, including: S210, calculating the positional difference of the inter-view correlation feature vector and the global correlation feature vector to obtain a differential feature vector; S220, calculating the positional sum of the inter-view correlation feature vector and the global correlation feature vector to obtain a sum feature vector; S230, calculating the full-scene homography correlation matrix between the inter-view correlation feature vector and the global correlation feature vector; S240, performing a logarithmic operation on the sum feature vector to obtain a logarithmic sum feature vector, wherein the logarithmic operation on the sum feature vector is The operation represents calculating the logarithmic function value of the eigenvalue of each position in the summed eigenvector; S250, calculating the norm of the differential eigenvector as the hierarchical depth characteristic value of the differential eigenvector; S260, calculating the Frobenius norm of the full-scene homography association matrix as the depth perception value of the full-scene homography association matrix; S270, using the hierarchical depth characteristic value as a weighting coefficient to weight the logarithmic summed eigenvector by position to obtain a weighted eigenvector; and, S280, using the depth perception value as a bias, calculating the sum of the depth perception value and the eigenvalue of each position in the weighted eigenvector to obtain the classified eigenvector.

[0080] In step S180, the classification feature vector is passed through a classifier to obtain a classification result, which is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets the predetermined requirements. That is, in a specific example, the classification feature vector is processed by the classifier using the following formula to obtain the classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|X}, where W1 to W n is the weight matrix, B1 to B n is the bias vector, and X is the classification feature vector.

[0081] In summary, the intelligent manufacturing method of the stainless steel cabinet of the embodiment of the present application is explained, which adopts a convolutional neural network model based on deep learning technology to extract implicit correlation features from multiple perspective images of the assembled stainless steel cabinet collected from multiple perspectives from both local and global aspects, and uses a layered deep homography alignment fusion method to perform homography alignment of the layered scene depth stream based on vector differential expression during feature fusion, so as to improve the classification effect of the fused features, and thus can accurately detect the assembly quality of the assembled stainless steel cabinet.

[0082] Exemplary Systems

[0083] Figure 5 FIG2 is a block diagram of an intelligent manufacturing system for a stainless steel cabinet according to an embodiment of the present application. Figure 5 As shown, the intelligent manufacturing system 500 of the stainless steel cabinet according to the embodiment of the present application includes: an image data acquisition unit 510, which is used to collect multiple perspective images of the assembled stainless steel cabinet from multiple perspectives through a camera; a local feature extraction unit 520, which is used to pass each perspective image of the multiple perspective images through a first convolutional neural network as a feature extractor to obtain multiple local perspective feature maps; a three-dimensional convolution unit 530, which is used to arrange the multiple local perspective feature maps into a three-dimensional input tensor and then obtain an inter-perspective correlation feature map through a second convolutional neural network using a three-dimensional convolution kernel; a panoramic stitching unit 540, which is used to perform panoramic stitching on the multiple local perspective feature maps. After splicing, a panoramic feature map is obtained; a global feature extraction unit 550 is used to pass the panoramic feature map through a non-local neural network to obtain a global correlation feature map; a dimensionality reduction unit 560 is used to expand the inter-view correlation feature map and the global correlation feature map into vectors respectively to obtain an inter-view correlation feature vector and a global correlation feature vector; a feature fusion unit 570 is used to fuse the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector; and a classification unit 580 is used to pass the classification feature vector through a classifier to obtain a classification result, and the classification result is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets the predetermined requirements.

[0084] In one example, in the above-mentioned intelligent manufacturing system 500 for stainless steel cabinets, the local feature extraction unit 520 is further used to: each layer of the first convolutional neural network performs the following on the input data in the forward pass of the layer: convolution processing on the input data to obtain a convolution feature map; pooling processing on the convolution feature map to obtain a pooling feature map; and nonlinear activation on the pooling feature map to obtain an activation feature map; wherein the output of the last layer of the first convolutional neural network is the multiple local perspective feature maps, and the input of the first layer of the first convolutional neural network is each perspective image in the multiple perspective images.

[0085] In one example, in the above-mentioned intelligent manufacturing system 500 for stainless steel cabinets, the three-dimensional convolution unit 530 is further used to: the second convolutional neural network using the three-dimensional convolution kernel performs the following on the input data in the forward pass of the layer: performing three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; performing pooling processing on the convolution feature map to obtain a pooled feature map; and performing nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the second convolutional neural network is the inter-view correlation feature map, and the input of the first layer of the second convolutional neural network is the three-dimensional input tensor.

[0086] In one example, in the intelligent manufacturing system 500 for the stainless steel cabinet, the panoramic stitching unit 540 is further configured to stitch the multiple local perspective feature maps along a height dimension or a width dimension to obtain the panoramic feature map.

[0087] In one example, in the above-mentioned intelligent manufacturing system 500 for stainless steel cabinets, the global feature extraction unit 550 is further used to: input the panoramic feature map into the first point convolution layer, the second point convolution layer and the third point convolution layer of the non-local neural network respectively to obtain a first feature map, a second feature map and a third feature map; calculate the position-weighted sum of the first feature map and the second feature map to obtain an intermediate fusion feature map; input the intermediate fusion feature map into the Softmax function to normalize the feature values of each position in the intermediate fusion feature map to obtain a normalized intermediate fusion feature map; calculate the position-weighted sum of the normalized intermediate fusion feature map and the third feature map to obtain a re-fusion feature map; embed the re-fusion feature map through a Gaussian similarity function to calculate the similarity between the feature values of each position in the re-fusion feature map to obtain a global perception feature map; pass the global perception feature map through the fourth point convolution layer of the non-local neural network to obtain a channel-adjusted global perception feature map; and calculate the position-weighted sum of the channel-adjusted global perception feature map and the high-dimensional correlation local feature map to obtain the global correlation feature map.

[0088] In one example, in the above-mentioned intelligent manufacturing system 500 for stainless steel cabinets, the dimensionality reduction unit 560 is further used to: divide each feature matrix in the inter-view correlation feature map and each feature matrix in the global correlation feature map with row vectors as division units to obtain multiple first row vectors corresponding to the inter-view correlation feature map and multiple second row vectors corresponding to the global correlation feature map; splice the multiple first row vectors to obtain the inter-view correlation feature vector; and splice the multiple second row vectors to obtain the global correlation feature vector.

[0089] In one example, in the above-mentioned intelligent manufacturing system 500 for stainless steel cabinets, Figure 6 As shown, the feature fusion unit 570 includes: a differential subunit 571 for calculating the positional difference between the inter-view correlation feature vector and the global correlation feature vector to obtain a differential feature vector; a summation subunit 572 for calculating the positional summation of the inter-view correlation feature vector and the global correlation feature vector to obtain a summation feature vector; an association matrix subunit 573 for calculating the full-scene homography association matrix between the inter-view correlation feature vector and the global correlation feature vector; a logarithm calculation subunit 574 for performing a logarithm operation on the summation feature vector to obtain a logarithmic summation feature vector, wherein the logarithm operation on the summation feature vector represents the calculation of each of the summation feature vectors. The logarithmic function value of the eigenvalue of the position; the eigenvalue calculation subunit 575, used to calculate the norm of the differential eigenvector as the hierarchical depth characteristic value of the differential eigenvector; the depth perception value calculation subunit 576, used to calculate the Frobenius norm of the full-scene homography association matrix as the depth perception value of the full-scene homography association matrix; the weighting subunit 577, used to weight the logarithmic sum eigenvector by position using the hierarchical depth characteristic value as a weighting coefficient to obtain a weighted eigenvector; and the fusion subunit 578, used to calculate the sum of the depth perception value and the eigenvalue of each position in the weighted eigenvector using the depth perception value as a bias to obtain the classification eigenvector.

[0090] In one example, in the intelligent manufacturing system 500 for the stainless steel cabinet, the association matrix subunit is further used to calculate the product between the transposed vector of the inter-view association feature vector and the global association feature vector to obtain the full-scene homography association matrix.

[0091] In one example, in the intelligent manufacturing system 500 for stainless steel cabinets, the classification unit 580 is further configured to: use the classifier to process the classification feature vector using the following formula to obtain the classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|X}, where W1 to W n is the weight matrix, B1 to B n is the bias vector, and X is the classification feature vector.

[0092] Here, those skilled in the art will appreciate that the specific functions and operations of the various units and modules in the intelligent manufacturing system 500 for stainless steel cabinets have been described in detail above. Figures 1 to 4The description of the intelligent manufacturing method of the stainless steel cabinet has been introduced in detail, and therefore, its repeated description will be omitted.

[0093] As described above, the intelligent manufacturing system 500 for stainless steel cabinets according to the embodiment of the present application can be implemented in various terminal devices, such as a server for the intelligent manufacturing algorithm for stainless steel cabinets. In one example, the intelligent manufacturing system 500 for stainless steel cabinets according to the embodiment of the present application can be integrated into the terminal device as a software module and / or hardware module. For example, the intelligent manufacturing system 500 for stainless steel cabinets can be a software module in the operating system of the terminal device, or can be an application developed for the terminal device; of course, the intelligent manufacturing system 500 for stainless steel cabinets can also be one of the many hardware modules of the terminal device.

[0094] Alternatively, in another example, the intelligent manufacturing system 500 of the stainless steel cabinet and the terminal device may also be separate devices, and the intelligent manufacturing system 500 of the stainless steel cabinet may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

Claims

1. An intelligent manufacturing method for stainless steel cabinets, characterized in that: include: The camera collects multiple perspective images of the assembled stainless steel cabinet from multiple perspectives; Passing each of the multiple perspective images through a first convolutional neural network as a feature extractor to obtain multiple local perspective feature maps; Arranging the multiple local view feature maps into a three-dimensional input tensor and obtaining an inter-view correlation feature map through a second convolutional neural network using a three-dimensional convolution kernel; Performing panoramic stitching on the multiple local view feature maps to obtain a panoramic feature map; Passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map; Expanding the inter-view correlation feature map and the global correlation feature map into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector respectively; fusing the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector; as well as Passing the classification feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements; Wherein, fusing the inter-perspective correlation feature vector and the global correlation feature vector to obtain a classification feature vector includes: calculating the positional difference between the inter-perspective correlation feature vector and the global correlation feature vector to obtain a differential feature vector; calculating the positional sum of the inter-perspective correlation feature vector and the global correlation feature vector to obtain a sum feature vector; calculating the full-scene homography correlation matrix between the inter-perspective correlation feature vector and the global correlation feature vector; performing a logarithmic operation on the sum feature vector to obtain a logarithmic sum feature vector, wherein the logarithmic operation table of the sum feature vector is The method comprises the following steps: calculating the logarithmic function value of the eigenvalue of each position in the summed eigenvector; calculating the -norm of the differential eigenvector as the hierarchical depth characteristic value of the differential eigenvector; calculating the Frobenius norm of the full-scene homography association matrix as the depth perception value of the full-scene homography association matrix; weighting the logarithmic summed eigenvector by position using the hierarchical depth characteristic value as a weighting coefficient to obtain a weighted eigenvector; and calculating the sum of the depth perception value and the eigenvalue of each position in the weighted eigenvector using the depth perception value as a bias to obtain the classification eigenvector.

2. The intelligent manufacturing method of stainless steel cabinets according to claim 1, wherein: Passing each of the plurality of perspective images through a first convolutional neural network as a feature extractor to obtain a plurality of local perspective feature maps, including: performing the following on the input data in the forward pass of each layer of the first convolutional neural network: Perform convolution on the input data to obtain a convolution feature map; Performing pooling processing on the convolution feature map to obtain a pooled feature map; and Performing nonlinear activation on the pooled feature map to obtain an activated feature map; Among them, the output of the last layer of the first convolutional neural network is the multiple local perspective feature maps, and the input of the first layer of the first convolutional neural network is each perspective image in the multiple perspective images.

3. The intelligent manufacturing method of stainless steel cabinet according to claim 2, wherein: Arranging the multiple local view feature maps into a three-dimensional input tensor and obtaining an inter-view correlation feature map by using a second convolutional neural network with a three-dimensional convolution kernel includes: performing the following operations on the input data in a forward pass of the layer by the second convolutional neural network with the three-dimensional convolution kernel: Performing three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; Performing pooling processing on the convolution feature map to obtain a pooled feature map; and Performing nonlinear activation on the pooled feature map to obtain an activated feature map; The output of the last layer of the second convolutional neural network is the inter-view correlation feature map, and the input of the first layer of the second convolutional neural network is the three-dimensional input tensor.

4. The intelligent manufacturing method of stainless steel cabinets according to claim 3, wherein: The multiple local view feature maps are panorama-stitched to obtain a panoramic feature map, including: The multiple local perspective feature maps are spliced along the height dimension or the width dimension to obtain the panoramic feature map.

5. The intelligent manufacturing method of stainless steel cabinets according to claim 4, wherein: Passing the panoramic feature map through a non-local neural network to obtain a global correlation feature map includes: Inputting the panoramic feature map into the first point convolution layer, the second point convolution layer and the third point convolution layer of the non-local neural network respectively to obtain a first feature map, a second feature map and a third feature map; Calculating a position-weighted sum of the first feature map and the second feature map to obtain an intermediate fused feature map; Inputting the intermediate fusion feature map into a Softmax function to normalize the feature values of each position in the intermediate fusion feature map to obtain a normalized intermediate fusion feature map; Calculating a position-weighted sum of the normalized intermediate fusion feature map and the third feature map to obtain a re-fusion feature map; The re-fused feature map is embedded in a Gaussian similarity function to calculate the similarity between the feature values at each position in the re-fused feature map to obtain a global perception feature map; Passing the global perception feature map through the fourth point convolution layer of the non-local neural network to obtain a channel-adjusted global perception feature map; and A position-weighted sum of the channel-adjusted global perceptual feature map and the high-dimensional correlated local feature map is calculated to obtain the global correlated feature map.

6. The intelligent manufacturing method of stainless steel cabinets according to claim 5, wherein: Expanding the inter-view correlation feature map and the global correlation feature map into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector, respectively, includes: Dividing each feature matrix in the inter-view correlation feature map and each feature matrix in the global correlation feature map using a row vector as a division unit to obtain a plurality of first row vectors corresponding to the inter-view correlation feature map and a plurality of second row vectors corresponding to the global correlation feature map; splicing the plurality of first row vectors to obtain the inter-view correlation feature vector; and The multiple second row vectors are concatenated to obtain the global correlation feature vector.

7. The intelligent manufacturing method of stainless steel cabinets according to claim 6, wherein: Calculating a full-scene homography correlation matrix between the inter-view correlation feature vector and the global correlation feature vector, comprising: The product of the transposed vector of the inter-view correlation feature vector and the global correlation feature vector is calculated to obtain the full-scene homography correlation matrix.

8. The intelligent manufacturing method of stainless steel cabinets according to claim 7, wherein: Passing the classification feature vector through a classifier to obtain a classification result includes: The classification feature vector is processed using the classifier using the following formula to obtain the classification result, wherein the formula is: softmax{(W n ,B n ):…:(W1,B1)|X}, where W1 to W n is the weight matrix, B1 to B n is the bias vector, and X is the classification feature vector.

9. An intelligent manufacturing system for stainless steel cabinets, characterized in that: include: An image data acquisition unit, configured to capture images of the assembled stainless steel cabinet from multiple viewing angles using a camera; a local feature extraction unit, configured to pass each of the plurality of perspective images through a first convolutional neural network serving as a feature extractor to obtain a plurality of local perspective feature maps; A three-dimensional convolution unit is configured to arrange the multiple local view feature maps into a three-dimensional input tensor and obtain an inter-view correlation feature map by using a second convolutional neural network with a three-dimensional convolution kernel; a panoramic stitching unit, configured to perform panoramic stitching on the multiple local view feature maps to obtain a panoramic feature map; A global feature extraction unit, configured to pass the panoramic feature map through a non-local neural network to obtain a global correlation feature map; a dimensionality reduction unit, configured to expand the inter-view correlation feature map and the global correlation feature map into vectors to obtain an inter-view correlation feature vector and a global correlation feature vector; a feature fusion unit, configured to fuse the inter-view correlation feature vector and the global correlation feature vector to obtain a classification feature vector; as well as a classification unit, configured to pass the classification feature vector through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the assembly quality of the assembled stainless steel cabinet meets predetermined requirements; Wherein, fusing the inter-perspective correlation feature vector and the global correlation feature vector to obtain a classification feature vector includes: calculating the positional difference between the inter-perspective correlation feature vector and the global correlation feature vector to obtain a differential feature vector; calculating the positional sum of the inter-perspective correlation feature vector and the global correlation feature vector to obtain a sum feature vector; calculating the full-scene homography correlation matrix between the inter-perspective correlation feature vector and the global correlation feature vector; performing a logarithmic operation on the sum feature vector to obtain a logarithmic sum feature vector, wherein the logarithmic operation table of the sum feature vector is The method comprises the following steps: calculating the logarithmic function value of the eigenvalue of each position in the summed eigenvector; calculating the -norm of the differential eigenvector as the hierarchical depth characteristic value of the differential eigenvector; calculating the Frobenius norm of the full-scene homography association matrix as the depth perception value of the full-scene homography association matrix; weighting the logarithmic summed eigenvector by position using the hierarchical depth characteristic value as a weighting coefficient to obtain a weighted eigenvector; and calculating the sum of the depth perception value and the eigenvalue of each position in the weighted eigenvector using the depth perception value as a bias to obtain the classification eigenvector.

Images