A deep learning-based copper alloy component quality detection method

Abstract

The application provides a kind of copper alloy element quality detection method based on deep learning, S1: through laser scanning, the acquisition of initial point cloud data of copper alloy element is carried out;S2: using bilateral filtering algorithm, the collected data is preprocessed, and the point cloud data of copper alloy element after denoising is obtained;S3: based on deep learning, the size detection classification of copper alloy element is carried out, and the size detection result is obtained;S4: the copper alloy element qualified in size enters surface defect detection, in surface defect detection, the YOLOv7 network model is improved, the attention mechanism is increased while being lightened, and the detection result is obtained based on the improved YOLOv7 network model.The application uses deep learning to detect the surface quality detection and size quality detection of copper alloy element.In the aspect of surface quality detection, targeted improvement is carried out on the original YOLOv7, so that the network can be more suitable for small size and large scale change defects, and is more easily deployed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of copper alloy element quality detection, and particularly relates to a copper alloy element quality detection method based on deep learning. BACKGROUND

[0002] Copper alloy is a key material for producing heat dissipation elements due to its outstanding heat conduction properties, but its production process is complex, and there are many factors affecting the quality of copper alloy elements, which may cause defects in the appearance of the copper alloy element surface, the size of the element exceeds the specified quality problems, thereby affecting the outgoing rate of the copper alloy element and causing economic losses. At present, copper alloy processing still relies on manual visual inspection, which is low in efficiency and large in error. According to research, when the target defect size is less than 0.5mm or the optical deformation is small, the artificial visual method cannot detect the defect. In addition, when manually detecting the size, a micrometer or other measuring tools are often used for measurement. In this process, the accuracy of the measuring tools and the subjectivity of the person will greatly affect the measurement. One of the key tasks for realizing intelligent manufacturing in China is to accelerate the development of artificial intelligence technology in the application of quality detection in the manufacturing industry. Therefore, in order to reduce the error caused by manual work, it is necessary to introduce a more efficient and reliable detection method. With the rapid development of deep learning technology and its application in the industrial field, introducing deep learning into factory production is considered an important direction of industrial intelligence. SUMMARY

[0003] In view of the above problems of the prior art, the present application aims to provide a copper alloy element quality detection method based on deep learning, which meets the requirements of high accuracy, high precision, real-time performance and the like.

[0004] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: a copper alloy element quality detection method based on deep learning, comprising the following steps:

[0005] S1: collecting initial point cloud data of the copper alloy element by laser scanning;

[0006] S2: pre-processing the collected data using a bilateral filtering algorithm to obtain denoised copper alloy element point cloud data;

[0007] S3: performing size detection and classification of the copper alloy element based on deep learning to obtain a size detection result;

[0008] S4: the copper alloy element with qualified size enters the surface defect detection, and in the surface defect detection, the YOLOv7 network model is improved, the light weight processing is performed on the YOLOv7 network model while the attention mechanism is added, and the surface defect detection is performed based on the improved YOLOv7 network model to obtain a detection result.

[0009] Further, the bilateral filtering in S2 adopts the combination of two Gaussian filters, one responsible for calculating the weight of spatial proximity, and the other responsible for calculating the weight of pixel value similarity, under the simultaneous action of the two Gaussian filters, it is bilateral filtering, the expression is as follows:

[0010]

[0011] Where g(i,j) represents the output point; S(i,j) represents the range of (2N+1)(2N+1) size centered on (i,j); f(k,l) represents the input point; w(i,j,k,l) represents the value calculated by two Gaussian functions;

[0012] W=W s *W r ,w s is a spatial proximity Gaussian function, which measures the distance between p(i,j) and q(k,l), the farther the distance, the lower the weight, W r is a pixel value similarity Gaussian function, which measures the pixel value similarity between p and q, the more similar the weight is larger;

[0013]

[0014] Where f(i,j) represents the pixel value corresponding to p(i,j), σ s is the spatial domain standard deviation, σ r is the value domain standard deviation.

[0015] Further, in S3, the size detection and classification of the copper alloy element based on deep learning includes the following steps:

[0016] The PointNet network is used to classify the preprocessed point cloud data, the input points are transformed by a spatial transformation network, and then input into another multilayer perception after using a multilayer perception MLP and feature transformation, and the overall features are extracted by a maximum pooling operation, and then the classification result is obtained by a multilayer perception.

[0017] Further, in S4, a triple attention mechanism is introduced to improve the network structure, which includes the following three levels of attention calculation:

[0018] Global attention: global attention is calculated on the most bottom layer or the most original representation of the input data, at this level, the model tries to understand the context information of the entire input, and calculates the importance of each part;

[0019] Group attention: group attention is calculated on the basis of the representation obtained by global attention, which divides the global representation into different groups, and then calculates the attention weight between these groups;

[0020] Intra-group attention: Intra-group attention is calculated based on the representation obtained from inter-group attention. It calculates attention weights within each group to capture the importance and relevance of the parts within the group.

[0021] The Z-pool layer is responsible for reducing the C-dimensional tensor to 2-dimensionality, and connecting the average pooling feature and the maximum pooling feature in that dimension, expressed as follows:

[0022] Z-pool(X) = [MaxPool 0d (X),AvgPool 0d (X)] (5)

[0023] Where 0d is the 0th dimension in which max and average pooling operations occur;

[0024] Given an input tensor X∈R C×H×W First, it is passed to the three branches in the triple attention module;

[0025] In the first branch, an interaction is established between the C and W dimensions. The input tensor X is rotated 90° counterclockwise along the W axis. This rotation tensor... The shape represented is (H×C×W), and then the tensor after Z-Pool is... (2×C×W), then through a standard convolutional layer with a kernel size of k×k, followed by a batch normalization layer, to provide an intermediate output with a dimension of (1×C×W); then, the tensor is used to generate attention weights through a sigmoid function; finally, the output is rotated 90° clockwise along the W axis to maintain the same shape as the input.

[0026] In the second branch, an interaction is established between the H and C dimensions. The input tensor X is rotated 90° counterclockwise along the H axis. This rotated tensor... The shape is represented as (W×H×C), and then the tensor after Z-Pool is... The shape is (2×H×C), and then it is passed through a standard convolutional layer with a kernel size of k×k, followed by a batch normalization layer to provide an intermediate output with a dimension of (1×H×C); then, the tensor is used to generate attention weights through a sigmoid function; finally, the output is rotated 90° clockwise along the H axis to maintain the same shape as the input.

[0027] In the third branch, an interaction is established between the H dimension and the W dimension: the channels of the input tensor X are simplified to 2 by Z-pool, the simplified tensor of this shape (2xHxW) is then simplified by a standard convolution layer defined by a kernel size kxk, and then by a batch normalization layer; the output is generated by a sigmoid activation layer to generate a shape (1xHxW) attention weight, which is applied to the input X to obtain the result The fine tensors (CxHxW) generated by the three branches are then aggregated together by simple averaging;

[0028] The final output tensor is:

[0029]

[0030] Where σ represents the sigmod activation function; And represent the standard two-dimensional convolution layer defined by the kernel size k in the three branches of triple attention.

[0031] Further, MobileNetV3 is used in S4 to lightweight the network structure. MobileNet modifies the standard convolution and designs a depth separable convolution; the depth separable convolution structure is composed of two groups of convolutions, namely depth convolution and point convolution; the depth convolution uses a convolution kernel with a channel number of one, and each convolution kernel is responsible for a separate channel, that is, the number of convolution kernels is consistent with the input channel output; the point convolution kernel size is 1x1, which performs dimension lifting and dimension reduction on the feature map.

[0032] Depth convolution: the convolution kernel size of depth convolution is D K ×D K ×1, the number is M, and each needs to do D F ×D F times of multiplication and addition operation, and the calculation amount is:

[0033] D k ×D k ×D F ×D F ×M (7)

[0034] The convolution kernel size of point convolution is 1x1xM, the number is N, and each needs to do D F ×D F times of multiplication and addition operation, and the calculation amount is:

[0035] D F ×D F ×MxN (8)

[0036] The calculation amount of depth separable convolution is:

[0037] D kX D k X D F X D F X M+D F X D F X M X N (9)

[0038] The beneficial effects of the present application are:

[0039] The present application uses deep learning to detect the surface quality and size quality of copper alloy elements. In terms of surface quality detection, the present application is based on the actual application scene and the characteristics of the researched copper alloy surface defects, and the original YOLOv7 is improved in a targeted manner, so that the network can be more suitable for small size and large scale change defects. In order to meet the factory precision and real-time requirements, the network is lightened, which is easier to deploy. In terms of size quality detection, the present application designs a size quality detection framework based on deep learning, which directly learns the specification features of the copper alloy element point cloud model through deep learning method, compares with the features of the standard element, and realizes the specification judgment of the element. Since it is not necessary to measure specific data, the method of the present application greatly reduces the size quality detection time, and the present application also has high detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 is the overall flowchart of the present application;

[0041] Figure 2 is the structure diagram of the spatial transformation network;

[0042] Figure 3 is the structure diagram of the PointNet network;

[0043] Figure 4 is the branch model diagram of the triple attention mechanism;

[0044] Figure 5 is the comparison diagram of the depth separable convolution and the standard convolution;

[0045] Figure 6 is the specific form diagram of the H-swish activation function. DETAILED DESCRIPTION

[0046] The present application will be further described below in conjunction with specific embodiments, but the protection scope of the present application is not limited to this.

[0047] A kind of copper alloy component quality detection method based on deep learning, first, using laser scanning obtains the point cloud data of part, then using bilateral filtering removes the noise in data, avoid affecting the result of classification afterwards, using the point cloud model classification based on deep learning is carried out size quality detection, size quality qualified enters surface defect detection.In surface defect quality detection, the present application uses yolo as basic model, in order to adapt to the need of actual detection speed and high precision, the present application has made targeted improvement on yolo basic model, so that the whole model is more lightweight, while more efficient feature extraction.The overall flow chart is as shown in Figure 1 As shown in the following steps:

[0048] S1: the initial point cloud data of copper alloy component is collected by laser scanning;

[0049] S2: using bilateral filtering algorithm for data preprocessing, obtains the point cloud data of copper alloy component after denoising;

[0050] For the copper alloy component produced, there are many places that need manual inspection, such as length, width, thickness, fillet radius, etc.In order to improve the detection efficiency and save the detection time, the present application uses the method based on deep learning to extract the point cloud data features of standard component, and compares with the data features of the detected component to directly judge whether the component is qualified.

[0051] Due to the influence of scanning environment, noise points may be generated in the collected point cloud data, which may interfere with the subsequent detection, so the collected point cloud data needs to be preprocessed.The pre-processing operation adopts bilateral filtering algorithm.

[0052] Bilateral filtering is a kind of nonlinear filtering.It is a processing method combining the spatial proximity of image and the similarity of pixel value.During filtering, the filtering method considers the spatial proximity information and color similarity information at the same time, which can filter out noise and smooth image while preserving edges.Bilateral filtering uses the combination of two Gaussian filters.One is responsible for calculating the weight of spatial proximity, that is, the principle of commonly used Gaussian filter.The other is responsible for calculating the weight of pixel value similarity.Under the simultaneous action of the two Gaussian filters, it is bilateral filtering, and the expression is as follows:

[0053]

[0054] Where g (i,j) represents the output point;S (i,j) represents the range of (2N+1) (2N+1) size with (i,j) as the center;f (k,l) represents the input point;w (i,j,k,l) represents the value calculated by two Gaussian functions.

[0055] W=W s *Wr w s is the spatial proximity Gaussian function, which measures the distance between p(i,j) and q(k,l), the farther the distance, the lower the weight r is the pixel value similarity Gaussian function, which measures the pixel value similarity between p and q, the more similar the weight is greater;

[0056]

[0057] where f(i,j) represents the pixel value corresponding to p(i,j), σ s is the spatial domain standard deviation, σ r is the value domain standard deviation.

[0058] w is the product of w s and w r . For w s , it is a normal Gaussian filter function, the coordinates of which are substituted, and σ is a program input value, and the function is calculated in the spatial proximity. And w r is the pixel value similarity (color space) calculation, note that this is the Gaussian function substituted by the coordinate value, the norm above is ‖f(i,j)-f(k,l)‖ 2 , that is, the square of the absolute value of the difference between the pixel values of the two points.

[0059] Through step S2, all points in the point cloud are processed, and finally the denoised point cloud data is obtained.

[0060] S3: based on deep learning, the size of the copper alloy element is detected and classified, and the size detection result is obtained.

[0061] According to step S2, the element point cloud data after denoising is obtained, and in step S3, the denoised data needs to be input into the model for classification to obtain the result of whether the element size is qualified. The size quality detection proposed in the present application is also based on the 3D point cloud model classification idea based on deep learning of step S3. In the present application, the copper alloy element to be detected has a relatively simple shape, and there is a high demand for detection speed and accuracy, so the PointNet network with high calculation efficiency and good global feature extraction capability is selected to classify the preprocessed point cloud data.

[0062] For a point cloud, translation and rotation should still represent the same object. To solve this problem, the pointnet network used in the present application newly introduces a T-Net network to learn the rotation of the point cloud, calibrates the object, so that the subsequent network only needs to classify the calibrated point cloud information. For the rotation of three-dimensional points, only a 3x3 rotation matrix needs to be multiplied to correct it; similarly, after mapping the point cloud to a K-dimensional redundant space, a kxk rotation matrix is also learned, which optimizes the network's performance in classifying the point cloud, so that the network has invariance to the rotation and position of the input point cloud. In PointNet, in order to solve the problem of rotation invariance of point cloud data, two spatial transformation networks are introduced. The first input transformation is to adjust the point cloud in space to rotate to a more advantageous angle for classification, such as turning the object to the front; the second feature conversion is to align the extracted 64-dimensional features, that is, to transform the point cloud at the feature level. The structure is as shown in Figure 2 .

[0063] Point cloud is essentially a long string of points (nx3 matrix, where n is the number of points). Geometrically, the order of points does not affect its representation of the overall shape in space. In PointNet, after the network extracts certain features from each point, it can extract global features from the entire point cloud using the maximum pooling operation.

[0064] The overall process of size detection and classification of copper alloy elements based on deep learning is as follows: the input points are subjected to spatial transformation network, then subjected to multi-layer perception (MLP) and feature transformation, and then input into another multi-layer perception to extract overall features through maximum pooling operation, and then obtain classification results through multi-layer perception. The structure of the PointNet network is as shown in Figure 3 .

[0065] S4: The size qualified copper alloy element enters the surface defect detection, in the surface defect detection, the YOLOv7 network model is improved, the attention mechanism is increased while the YOLOv7 network model is lightened, and the detection result is obtained based on the improved YOLOv7 network model.

[0066] Target detection is an important research direction in the field of computer vision. Traditional target detection methods need to perform object recognition and object positioning in two steps, which usually requires a large amount of computing resources and time. In recent years, deep learning-based target detection methods have been widely applied, among which the YOLO algorithm has become a target detection technology that has attracted much attention due to its speed and accuracy.

[0067] In view of the difficulty in detecting small-scale defects in actual production, the present application designs a model improved based on an attention mechanism.

[0068] In actual production, copper alloy components have several different defects on their surface, such as dents and scratches, and these defects vary in shape and size. This requires the network to be adaptable to defects of different sizes, and at the same time, to have a good ability to extract the features of tiny surface defects.

[0069] Building upon this, this invention introduces a triple attention mechanism to improve the network structure. The attention mechanism is used in the feature fusion module of the model, helping the model focus on feature extraction of surface defects. The triple attention mechanism typically includes the following three levels of attention computation:

[0070] Global attention: Global attention is typically computed at the lowest or most primitive representation of the input data. At this level, the model attempts to understand the contextual information of the entire input and compute the importance of each part.

[0071] Inter-group attention: Inter-group attention is calculated based on the representation obtained from global attention. It divides the global representation into different groups and then calculates attention weights between these groups. This level of attention helps the model better understand the relationships and interactions between different parts of the input data.

[0072] Intra-group attention: Intra-group attention is calculated based on the representation obtained from inter-group attention. It calculates attention weights within each group to capture the importance and relevance of parts within the group. This level of attention helps the model better understand the inherent structure and semantic information of each part within the group.

[0073] The Z-pool layer is responsible for reducing the C-dimensional tensor to 2-dimensionality by concatenating the average pooling feature and the maximum pooling feature in that dimension. This allows the layer to preserve a rich representation of the actual tensor while reducing its depth to make further computation less computationally intensive. This can be expressed as:

[0074] Z-pool(X) = [MaxPool 0d (X),AvgPool 0d (X)] (5)

[0075] Where 0d is the 0th dimension in which max and average pooling operations occur;

[0076] Given an input tensor X∈R C×H×W First, it is passed to the three branches in the triple attention module.

[0077] In the first branch, an interaction is established between the C and W dimensions. To achieve this, the input tensor X is rotated 90° counterclockwise along the W axis. This rotated tensor... The shape of the tensor represented is (HxCxW), and then the tensor after Z-Pool The shape of the tensor represented is (HxCxW), and then the tensor after Z-Pool

[0078] In the second branch, an interaction is established between the H dimension and the C dimension, in order to achieve this, the input tensor X is rotated 90° counterclockwise along the H axis. The rotated tensor The shape of the tensor represented is (HxCxW), and then the tensor after Z-Pool The shape of the tensor represented is (HxCxW), and then the tensor after Z-Pool

[0079] In the third branch, an interaction is established between the H dimension and the W dimension: the channels of the input tensor X are simplified to 2 by Z-pool. The simplified tensor of this shape (2xHxW) is simplified and then passed through a standard convolution layer defined by a kernel size kxk, and then passed through a batch normalization layer. The output generates attention weights of the shape (1xHxW) through a sigmoid activation layer, and applies it to the input X to obtain the result The fine-grained tensors (CxHxW) generated by the three branches are then aggregated together through simple averaging.

[0080] The structure diagram is as shown in Figure 4 The final output tensor is:

[0081]

[0082] Where σ represents the sigmod activation function; And represent the standard two-dimensional convolution layers defined by the kernel size k in the three branches of triple attention.

[0083] Based on the above improvements, in view of the fact that most factories in actual production are mobile or embedded devices, and the computing resources are relatively tight, the application designs a lightweight yolo model, which reduces the parameters and complexity of the model.

[0084] The new model designed by the application adopts MobileNetV3 to lighten the network structure, MobileNet modifies the standard convolution, and designs a depth separable convolution.

[0085] MobileNetV3 inherits the depth separable convolution structure of V1 version, and realizes the separation of convolution operation into two parts, namely depth convolution and point convolution, by the design.

[0086] The comparison between depth separable convolution and standard convolution is shown in the following table. Figure 5

[0087] The application decomposes the standard convolution kernel of the original network to form a single-channel mode.

[0088] Depth convolution: the convolution kernel size of depth convolution is D K ×D K ×1, the number is M, and each needs to do D F ×D F times of multiplication and addition operation, so the calculation amount is:

[0089] D k ×D k ×D F ×D F ×M (7)

[0090] The convolution kernel size of point convolution is 1x1xM, the number is N, and each needs to do D F ×D F times of multiplication and addition operation, so the calculation amount is:​

[0091] D F ×D F ×M×N (8)

[0092] Therefore, the calculation amount of the modified depth separable convolution of the application is:

[0093] D k ×D k ×D F ×D F ×M+D F ×D F ×M×N (9)

[0094] In the standard convolution, the size of the convolution kernel is D K ×D K ×M, and there are N in total, each of which needs to be operated D F ×D F times (assuming that the size of the output feature map is also D F ×D F ), so the calculation amount of the standard convolution is D K ×D K ×M×N×D F ×D F .

[0095] If a common 3x3 convolution kernel is used, the parameter amount and the calculation amount using the depth separable convolution are reduced to about one ninth of the original (N is large, and can be ignored), and the ratio of the calculation amount is as follows:

[0096]

[0097] Activation function optimization, in the traditional YOLOv7 network, the initial activation function is SiLU function, but the activation function has the risk of gradient disappearance problem, and the value range of the function is (0, 1), which may cause loss of information. According to the above situation, the application selects H-swish activation function to replace SiLU function. The H-swish activation function has lower bound, no upper bound, smooth, non-monotonic characteristics, which effectively improves the shortcomings of SiLU function, and can suppress the problem of gradient disappearance. The specific form of the H-swish activation function is as shown in Figure 6 .

[0098]

[0099] The common residual structure first performs a dimension reduction operation through point convolution, then uses a 3*3 convolution to extract features, and finally performs a dimension promotion operation through point convolution. In the reverse residual structure used in the application, the dimension promotion operation is first performed through point convolution, then the feature extraction is realized through 3*3 channel-by-channel convolution, at this time the common 3*3 convolution is not used, and finally the dimension reduction operation is realized through point convolution. In this process, the reverse residual structure changes the order of dimension promotion and reduction, and replaces the 3*3 standard convolution with the DW convolution. Using the reverse residual structure, the number of channels can be expanded, so that the requirement of the backbone network for computing resources is more easily met, and the model is beneficial to feature extraction.

[0100] The above description is only a preferred embodiment of the present application, which is used to illustrate the technical solutions of the present application, but not to limit the present application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, the modification and replacement of some technical features will not deviate from the protection scope of the present application. Therefore, the protection scope of the present application patent should be subject to the appended claims.

Claims

1. A deep learning-based copper alloy component quality detection method, characterized by, It comprises the following steps: S1: collecting initial point cloud data of the copper alloy element by laser scanning; S2: pre-processing the collected data using a bilateral filtering algorithm to obtain denoised copper alloy element point cloud data; S3: performing size detection and classification on the copper alloy element based on deep learning to obtain a size detection result; S4: the copper alloy element that meets the size requirement enters the surface defect detection, in which the YOLOv7 network model is improved, the network structure is lightened, and an attention mechanism is added, and the improved YOLOv7 network model is used for surface defect detection to obtain a detection result; In S2, the bilateral filtering uses the combination of two Gaussian filters, one responsible for calculating the spatial proximity weight, and the other responsible for calculating the pixel value similarity weight. Under the simultaneous action of the two Gaussian filters, it is bilateral filtering, and the expression is as follows: ； Where g(i, j) represents the output point; S(i, j) represents the range of (2R+1)(2R+1) centered at (i,j); f(k, l) represents the input point; and w(i, j, k, l) represents the value calculated by the two Gaussian functions; For the spatial proximity Gaussian function, the distance between p(i,j) and q(k,l) is measured, and the farther the distance, the lower the weight, For the pixel value similarity Gaussian function, the pixel value similarity between p and q is measured, and the more similar the pixel value, the greater the weight. ； ； where f(i,j) represents the pixel value corresponding to p(i,j), s for spatial domain standard deviation, r for value range standard deviation; In S4, MobileNetV3 is used to lighten the network structure, MobileNet modifies the standard convolution and designs a depth separable convolution; the depth separable convolution structure is composed of two groups of convolution, namely depth convolution and point convolution; the depth convolution uses a convolution kernel with a channel number of one, and each convolution kernel is responsible for a single channel, that is, the number of convolution kernels is consistent with the input channel; the point convolution kernel size is 1x1, which performs dimension lifting and dimension reduction on the feature map; Depthwise convolution: the size of the convolution kernel D K ×D K ×1, M in number, each of which needs to be done D F ×D F times multiply and add operations, the amount of calculation is: ； The convolution kernel size of point convolution is 1 × 1 × M, and the number is N, and each needs to be done D F × D F times of multiplication and addition operation, and the calculation amount is: ； The calculation amount of the depth separable convolution is: 。 2.The copper alloy component quality detection method based on deep learning according to claim 1, characterized in that, In S3, the size detection and classification of the copper alloy element based on deep learning comprises the following steps: PointNet network is used to classify the pre-processed point cloud data, the input points are subjected to spatial transformation network, and after using multi-layer perception MLP and feature transformation, they are input into another multi-layer perception to extract overall features through maximum pooling operation, and then the classification result is obtained through multi-layer perception. 3.The deep learning-based copper alloy component quality detection method according to claim 2, characterized by, In S4, a triple attention mechanism is introduced to improve the network structure, which includes the following three levels of attention calculation: Global attention: global attention is calculated on the most bottom layer or the most original representation of the input data. In this level, the model tries to understand the context information of the entire input and calculates the importance of each part; Inter-group attention: inter-group attention is calculated on the basis of the representation obtained by global attention. It divides the global representation into different groups and calculates the attention weight between these groups; Intra-group attention: intra-group attention is calculated on the basis of the representation obtained by inter-group attention. It calculates the attention weight within each group to capture the importance and relevance of the parts within the group; The Z-pool layer is responsible for reducing the C-dimensional tensor to 2 dimensions, connecting the average pooled features and the maximum pooled features on this dimension, and using the following formula: ； wherein and is the 0th dimension of the operation where maximum and average pooling occurs; Given an input tensor X It is first passed into the three branches in the triple attention module; In the first branch, an interaction is established between the C and W dimensions, the input tensor X is rotated 90° counterclockwise along the W axis, this rotated tensor has shape H × C × W, then after Z-Pool the tensor has shape 2 × C × W, then, by a standard convolutional layer with kernel size × , then by a batch normalization layer, an intermediate output with dimensions 1 × C × W is provided; then the tensor passes through a sigmoid to generate the attention weight; finally, the output is rotated 90° clockwise along the W axis to keep the same shape as the input. In the second branch, an interaction is established between the H and C dimensions. The input tensor X is rotated 90° counterclockwise along the H axis. This rotation tensor... The shape represented is W × H × C, and then the tensor after Z-Pool is... The shape is 2×H×C, and then, through a kernel size of × The standard convolutional layer is followed by a batch normalization layer to provide an intermediate output of dimension 1×H×C; then, the tensor is used to generate attention weights through a sigmoid function; finally, the output is rotated 90° clockwise along the H-axis to maintain the shape of the input. In the third branch, an interaction is established between the H and W dimensions: the channels of the input tensor X are reduced by Z-pool to 2, and this reduced tensor of shape 2 x H x W is then reduced by a kernel of size × defined standard convolutional layer, followed by a batch normalization layer; the output is passed through a sigmoid activation layer to generate attention weights of shape 1 x H x W, which are applied to the input X to obtain the result ; the fine tensors C x H x W produced by the three branches are then aggregated together by a simple average. The final output tensor is: ； wherein represents a sigmod activation function; , and represents a standard two-dimensional convolutional layer defined by the kernel size in three branches of the triple attention.

Images