Crop disease identification method based on adaptive dynamic grouping

Abstract

The invention discloses a crop disease identification method based on adaptive dynamic grouping, and belongs to the technical field of crop disease identification, and the method comprises the steps: obtaining a crop disease image data set, and carrying out the data enhancement, and obtaining training set data; based on DGC dynamic grouping convolution and MLCS multi-level channel rearrangement, a crop disease recognition model is established, training is conducted through training set data, a trained crop disease recognition model is obtained, and crop disease recognition is completed through the trained crop disease recognition model. By combining the dynamic grouping and multi-level channel rearrangement technology, lightweight design of the crop disease recognition model is realized, the calculation complexity and parameter quantity of the neural network model can be effectively reduced, the training time is remarkably shortened, and the training efficiency is improved. And meanwhile, efficient and accurate disease detection can be realized on a mobile terminal or embedded equipment with limited computing resources.

Description

Technical Field

[0001] The invention belongs to the technical field of crop disease identification, and in particular relates to a crop disease identification method based on adaptive dynamic grouping. Background Art

[0002] In smart agriculture, crop disease identification is one of the most critical tasks, which plays a vital role in ensuring food security and improving agricultural productivity. Traditional manual identification methods are time-consuming and labor-intensive, and the accuracy of identification is often affected by the subjective judgment of the observer. Inspired by the superiority of deep neural networks, various deep learning models have been used for crop disease identification tasks, such as convolutional neural networks, which have achieved remarkable results in recognition accuracy by capturing local and global features. However, due to the increase in network depth and slow inference speed, these models require a lot of computing resources and storage space, making them difficult to deploy efficiently on mobile or embedded devices with limited computing power.

[0003] To solve this problem, lightweight neural networks came into being for crop disease recognition tasks on mobile devices. For example, the MobileNet series significantly reduced computational costs by using depthwise separable convolutions; EfficientNet-B0 adopted a compound scaling strategy to achieve a balance between different network depths, widths, and input resolutions, and showed strong performance in wheat disease recognition tasks; SqueezeNet significantly reduced model parameters by introducing the Fire module, and increased the number of output channels by adjusting the ratio of 1×1 and 3×3 convolutions to identify a variety of leaf diseases. However, how to improve model recognition efficiency while maintaining model accuracy remains a challenge. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a crop disease identification method based on adaptive dynamic grouping, which solves the problem of difficulty in improving the model identification efficiency while maintaining the accuracy of the crop disease identification model.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a crop disease identification method based on adaptive dynamic grouping, comprising the following steps:

[0006] S1: Obtain a crop disease image dataset and perform data enhancement to obtain training set data;

[0007] S2: Establish a crop disease recognition model based on DGC dynamic group convolution and MLCS multi-level channel rearrangement;

[0008] S3: According to the crop disease recognition model, the training set data is used for training to obtain a trained crop disease recognition model;

[0009] S4: According to the trained crop disease recognition model, the crop images are processed to complete the crop disease recognition.

[0010] The beneficial effects of the present invention are as follows: the present invention realizes a lightweight design of a crop disease recognition model by combining dynamic grouping and multi-level channel rearrangement technology, which can effectively reduce the computational complexity and parameter amount of the neural network model, significantly shorten the training time, and at the same time can realize efficient and accurate disease detection on mobile terminals or embedded devices with limited computing resources.

[0011] Further: the crop disease recognition model includes an inverse residual module, a first dynamic group recognition module, a second dynamic group recognition module and a classifier connected in sequence;

[0012] The input training set data of the inverse residual module is fused with the output data of the residual module through a skip connection and used as the input data of the first dynamic group recognition module;

[0013] The input data of the first dynamic group identification module is merged with the output data of the first dynamic group identification module through a jump connection and used as the input data of the second dynamic group identification module;

[0014] The input data of the second dynamic group identification module is fused with the output data of the second dynamic group identification module through a skip connection and serves as the input data of the classifier.

[0015] The beneficial effect of the above further scheme is that through the cascade and jump connection between modules, the feature extraction and generalization capabilities of the crop disease recognition model can be enhanced, the input features can be fully utilized, and the disease recognition capability can be improved.

[0016] Further: the first dynamic group identification module and the second dynamic group identification module have the same structure, both comprising a dynamic group convolution submodule, a conditional batch normalization submodule and a multi-level channel rearrangement submodule connected in sequence;

[0017] The inverse residual module is used to extract the features of the crop disease image to obtain a feature map;

[0018] The dynamic group convolution submodule is used to adaptively adjust the number of groups of the group convolution according to the feature map, and perform a convolution operation to obtain a feature map after dynamic group convolution;

[0019] The conditional batch normalization submodule is used to normalize the feature map after convolution;

[0020] The multi-level channel rearrangement submodule is used to perform multi-level channel rearrangement on the feature map after dynamic grouping convolution, and output the feature map of multi-level channel rearrangement;

[0021] The classifier is used to integrate the feature maps of the multi-level channel re-shots input into the classifier and output the crop disease recognition results.

[0022] The beneficial effects of the above further scheme are: through the inverse residual module and the dynamic grouping recognition module, adaptive convolution feature extraction can be achieved, and the data is normalized through the conditional batch normalization submodule, and the multi-level channel rearrangement submodule optimizes feature fusion, so that the crop disease recognition model can perform fast and accurate crop disease recognition.

[0023] Further: the inverse residual module includes a 1x1 first convolutional layer, a 3x3 depth-separable convolution, and a 1x1 second convolutional layer connected in sequence;

[0024] The 1x1 first convolutional layer is used to expand the number of input channels to a high dimension;

[0025] The 3x3 depthwise separable convolution is used to perform a convolution operation on each channel and extract spatial information;

[0026] The 1x1 second convolutional layer is used to compress the number of channels back to the dimension of the initial input channels.

[0027] The beneficial effects of the above further scheme are as follows: the inverse residual module expands the feature dimension through the 1x1 first convolutional layer, extracts spatial information through the 3x3 depthwise separable convolution, and compresses the number of channels through the 1x1 second convolutional layer, which can improve the depth and breadth of feature extraction, reduce the computational cost and the number of parameters, and enhance the generalization ability of the crop disease recognition model.

[0028] Further: the dynamic group convolution submodule adaptively adjusts the number of group convolutions, and the specific expression of the method is as follows:

[0029]

[0030] Among them, X is the input feature map, G(X) is the number of groups of group convolution, G min is the minimum number of groups, G max is the maximum number of groups, α is the adjustment coefficient, σ 2 (X) is the global variance of the input feature map, is the upper limit of the input feature variance, is the lower limit of the input feature variance, C in is the number of input channels, is the global variance of the cth channel, H is the height of the input feature map, W is the width of the input feature map, i and j are the indexes of the input feature map, c is the sequence number of the channel, μ c is the mean of the cth channel.

[0031] The dynamic group convolution submodule performs a dynamic group convolution operation to obtain a feature map after dynamic group convolution, and its specific expression is as follows:

[0032]

[0033] Among them, Y g (i,j,c g ) is the convolution operation of the g-th group convolution, k g (X) is the convolution kernel size of the g-th group convolution, is the number of input channels of the g-th group convolution, m and n are the internal indexes used to traverse the convolution kernel, c′ is the index of the input channel, i+m-1 and j+n-1 are the translation operations of the convolution kernel on the input feature map, and K g is the convolution kernel configuration of the g-th group convolution, c g is the index of the output channel of the g-th group convolution, k min is the minimum convolution kernel size, k max is the maximum convolution kernel size, σ 2 (X g ) is the variance of the g-th group convolution input features.

[0034] The above further beneficial effect is that by adaptively adjusting the number of groups according to the complexity of different features, redundant calculations can be reduced and the calculation efficiency of the crop disease recognition model can be improved.

[0035] Further: the feature map after the dynamic grouping convolution is subjected to multi-level channel rearrangement, and the specific steps of channel rearrangement are as follows:

[0036] A1: Perform a reshape operation based on the feature map after dynamic group convolution to obtain a matrix;

[0037] A2: Perform a transpose operation on the matrix to obtain a transposed matrix;

[0038] A3: Perform a flattening operation based on the transposed matrix to obtain a feature map with rearranged channels.

[0039] The expression of the feature map of the multi-level channel re-shoot is as follows:

[0040] Y″(i,j,c)=Shuffle(Concat(Y′ 1 (i,j,·),Y′ 2 (i,j,·),…,Y′G (i,j,·)))

[0041] =Shuffle(Concat(Shuffle(Y 1 (i,j,·)),…,Shuffle(Y G (i,j,·))))

[0042] Y′ g (i,j,c)=Shuffle(Y g (i,j,c g ))

[0043] Among them, Y″(i,j,c) is the feature map of multi-level channel re-shooting, Y′ g (i, j, c) is the feature map of the first-level channel rearrangement, Y g (i,j,c g ) is the feature map after dynamic group convolution, Shuffle is the channel reordering, and Concat is the connection operation.

[0044] The beneficial effect of the above further scheme is that the present invention can further improve the fusion of cross-group channel information of the crop disease recognition model through multi-level channel rearrangement, thereby enhancing the feature recognition ability of the crop disease recognition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a flow chart of a crop disease identification method based on adaptive dynamic grouping;

[0046] Figure 2 It is a flowchart of the dynamic grouping convolution submodule;

[0047] Figure 3 A schematic diagram of the process of channel rearrangement;

[0048] Figure 4 A schematic diagram of the process of the multi-level channel rearrangement submodule;

[0049] Figure 5 This is the overall framework diagram of the crop disease identification model. DETAILED DESCRIPTION

[0050] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0051] like Figure 1As shown, a crop disease identification method based on adaptive dynamic grouping includes the following steps:

[0052] S1: Obtain a crop disease image dataset and perform data enhancement to obtain training set data;

[0053] S2: Establish a crop disease recognition model based on DGC dynamic group convolution and MLCS multi-level channel rearrangement;

[0054] S3: According to the crop disease recognition model, the training set data is used for training to obtain a trained crop disease recognition model;

[0055] S4: According to the trained crop disease recognition model, the crop images are processed to complete the crop disease recognition.

[0056] In S1, a crop disease image dataset is obtained and data enhancement is performed to obtain training set data;

[0057] Data enhancement includes: image cropping, image flipping, image rotation, image displacement, image deformation, changing image brightness and other image operations; ensuring that no other data is added during the image enhancement process, so that it is more likely to generate image value under limited data conditions, effectively increase the amount of data and improve the generalization ability of the model; using random cropping to properly process the image, and fusing multiple cropped images to improve the experimental prediction effect and solve the image distortion problem caused by free scaling.

[0058] In one embodiment of the present invention, a unified hyperparameter setting and data enhancement strategy are used to process a crop disease image dataset, and the image is cropped to a size of 224x224, and data enhancement operations such as random horizontal flipping of the image are performed to enhance the perspective robustness of the model; finally, the image pixel values ​​are normalized to have a mean of 0 and a standard deviation of 1.

[0059] In S2, a crop disease recognition model is established based on DGC dynamic group convolution and MLCS multi-level channel rearrangement; the crop disease recognition model includes an inverse residual module, a first dynamic group recognition module, a second dynamic group recognition module and a classifier connected in sequence;

[0060] Among them, the input training set data of the inverse residual module is fused with the output data of the residual module through a jump connection, and is used as the input data of the first dynamic group identification module; the input data of the first dynamic group identification module is fused with the output data of the first dynamic group identification module through a jump connection, and is used as the input data of the second dynamic group identification module; the input data of the second dynamic group identification module is fused with the output data of the second dynamic group identification module through a jump connection, and is used as the input data of the classifier; through the jump connection, the useful information in the input data of each module can be fully utilized to guide the subsequent modules to perform more complex disease feature recognition, ensuring that the key feature information is effectively captured while the network model is lightweight.

[0061] The first dynamic group identification module and the second dynamic group identification module have the same structure, and both include a dynamic group convolution submodule, a conditional batch normalization submodule, and a multi-level channel rearrangement submodule connected in sequence;

[0062] The inverse residual module is used to extract the features of crop disease images and obtain feature maps;

[0063] The dynamic group convolution submodule is used to adaptively adjust the number of groups of the group convolution according to the feature map, and perform the convolution operation to obtain the feature map after the dynamic group convolution;

[0064] The conditional batch normalization submodule is used to normalize the feature map after convolution;

[0065] The multi-level channel rearrangement submodule is used to perform multi-level channel rearrangement on the feature map after dynamic grouping convolution and output the feature map of multi-level channel rearrangement;

[0066] The classifier is used to integrate the feature maps of the multi-level channel re-shots input into the classifier and output the crop disease recognition results.

[0067] In one embodiment of the present invention, the inverse residual module includes a 1x1 first convolutional layer, a 3x3 depth-separable convolution, a 1x1 second convolutional layer, and a skip connection connected in sequence;

[0068] The first 1x1 convolutional layer is used to expand the number of input channels to a high dimension to obtain richer feature information, and uses ReLu6 as the activation function;

[0069] 3x3 depthwise separable convolutions to perform convolution operations on each channel and extract spatial information while keeping computational cost low;

[0070] The 1x1 second convolutional layer is used to compress the number of channels back to the dimension of the initial input channel, which can reduce the number of parameters and improve computational efficiency.

[0071] In one embodiment of the present invention, Figure 2 As shown, it is a flow chart of the dynamic grouping convolution submodule. Different from the traditional fixed number of groups, the dynamic grouping convolution submodule of the present invention can measure the diversity of the input feature map according to its variance, and adaptively adjust the number of groups of the grouping convolution to perform the convolution operation;

[0072] Among them, the dynamic group convolution submodule adaptively adjusts the number of group convolutions according to the input feature map. Its specific expression is as follows:

[0073]

[0074] Among them, X is the input feature map, G(X) is the number of groups of group convolution, G min is the minimum number of groups, G max is the maximum number of groups, α is the adjustment coefficient, σ 2 (X) is the global variance of the input feature map, is the upper limit of the input feature variance, is the lower limit of the input feature variance, C in is the number of input channels, is the global variance of the cth channel, H is the height of the input feature map, W is the width of the input feature map, i and j are the indexes of the input feature map, c is the sequence number of the channel, μ c is the mean of the cth channel; when the variance of the input feature map is very small, using the minimum number of groups can reduce the amount of calculation; when the variance of the input feature map is very large, using the maximum number of groups can capture more detailed information in the feature map; the adjustment coefficient α can control the speed of change of the number of groups.

[0075] The dynamic group convolution submodule performs dynamic group convolution operation, and its specific expression is as follows:

[0076]

[0077] Among them, Y g (i,j,c g ) is the convolution operation of the g-th group convolution, k h (X) is the convolution kernel size of the g-th group convolution, is the number of input channels of the g-th group convolution, m and n are the internal indexes used to traverse the convolution kernel, c′ is the index of the input channel, i+m-1 and j+n-1 are the translation operations of the convolution kernel on the input feature map, and K g is the convolution kernel configuration for the g-th group convolution, c g is the index of the output channel of the g-th group convolution, k min is the minimum convolution kernel size, k maxis the maximum convolution kernel size, σ 2 (X g ) is the variance of the g-th group convolution input features.

[0078] In one embodiment of the present invention, the conditional batch normalization submodule can normalize the feature map after convolution and introduce a nonlinear activation function ReLU function to increase the nonlinear representation ability of the model.

[0079] In one embodiment of the present invention, a multi-level channel rearrangement submodule is used to perform multi-level channel rearrangement on the feature map after the dynamic convolution operation to increase the fusion capability of feature information; Figure 3 As shown in the figure, it is a schematic diagram of the process of channel rearrangement. Channel rearrangement is performed after dynamic group convolution. The feature map after dynamic group convolution is reshaped into a matrix with g rows and n columns, and the matrix is ​​transposed. The transposed matrix is ​​flattened to obtain the feature map of channel rearrangement, thereby realizing information fusion.

[0080] like Figure 4 As shown, it is a schematic diagram of the process of the multi-level channel rearrangement submodule, and its specific method is as follows:

[0081] The input feature map Divided into G groups, the output feature map is The output feature map is concatenated from the convolution results of each group. The output feature map Y after each group convolution is g The first-level channel rearrangement is obtained Among them, Y′ g (i,j,c)=Shuffle(Y g (i,j,·)) represents the feature map Y g Information of all channels at position (i, j);

[0082] After the first level of channel reordering, the feature maps Y = Concat(Y′ 1 ,Y′ 2 ,…,Y′ G ) and then rearrange the channels to obtain in

[0083] Y″(i,j,c)=Shuffle(Concat(Y′ 1 (i,j,·),Y′ 2 (i,j,·),…,Y′ G (i,j,·)))

[0084] =Shuffle(Concat(Shuffle(Y 1(i,j,·)),…,Shuffle(Y G (i,j,·))))#

[0085] Multi-level channel shuffling can improve feature fusion and enhance performance, but excessive channel reordering may lead to diminishing returns and increase computational complexity. It is possible to evaluate the performance of features after each channel reordering and determine whether to continue channel reordering based on the evaluated performance.

[0086] In S3, according to the crop disease recognition model, the training set data is used for training to obtain a trained crop disease recognition model;

[0087] In one embodiment of the present invention, during the training of the crop disease recognition model, a cross entropy loss function is used to measure the difference between the output label and the true label; the Adam optimizer is used to update the model parameters, the learning rate is set to 1.0e-3, the batch size during training is set to 16, the number of processes for data loading is set to 4, and the total number of training rounds is 100; in the verification and testing stages, the image size is first adjusted to 256x256, and a 224x224 image area is cropped at the center to ensure that the cropped image contains the core content of the original image.

[0088] In S4, the crop images are processed according to the trained crop disease recognition model to complete the crop disease recognition. Among them, the trained crop disease recognition model can realize real-time and accurate crop disease recognition, save it, and deploy it on low-computing power devices, and verify the stability and response speed of the model on the mobile terminal, so that it can quickly identify crop diseases in actual production scenarios.

[0089] The following is an explanation of the process of the crop disease identification model:

[0090] like Figure 5As shown in the figure, it is the overall framework structure diagram of the crop disease recognition model. The input of the crop disease recognition model is the image of the crop disease dataset. First, the image of the crop disease dataset is preliminarily processed by the inverse residual module IRB for feature extraction to generate a feature representation; then, the feature representation enters the first dynamic grouping recognition module Stage 1 after the connection operation C, and enters the conditional batch normalization module CBN after passing through a convolution layer Conv Layer; in the first dynamic grouping recognition module, it includes a dynamic grouping convolution submodule DGC and a multi-level channel rearrangement submodule MLCS; the dynamic grouping convolution submodule DGC adjusts the convolution calculation method of the feature through adaptive dynamic grouping, and the multi-level channel rearrangement submodule MLCS uses channel rearrangement technology to further enhance the feature expression ability; then, the feature representation is fused with the stage output through the skip connection Skip-connection and enters the second dynamic grouping recognition module Stage 2 and conditional batch normalization module CBN; in the second dynamic grouping recognition module, it includes a dynamic grouping convolution submodule DGC and a multi-level channel rearrangement submodule MLCS, which use different parameters from the first dynamic grouping recognition module, such as the number of groups N1 and N2 respectively; through dynamically adjusted convolution kernel size and channel grouping, the second dynamic grouping recognition module determines whether to perform channel rearrangement again through the features output by the first dynamic grouping recognition module, and extracts and integrates features; finally, the model output generates a prediction result Output through the classifier Classifier, thereby realizing the recognition of crop disease categories.

[0091] The beneficial effects of the present invention are as follows: the present invention realizes a lightweight design of a crop disease recognition model by combining dynamic grouping and multi-level channel rearrangement technology, which can effectively reduce the computational complexity and parameter amount of the neural network model, significantly shorten the training time, and at the same time can realize efficient and accurate disease detection on mobile terminals or embedded devices with limited computing resources.

Claims

1. A crop disease identification method based on adaptive dynamic grouping, characterized in that: The following steps are involved: S1: Obtain a crop disease image dataset and perform data enhancement to obtain training set data; S2: Establish a crop disease recognition model based on DGC dynamic group convolution and MLCS multi-level channel rearrangement; S3: According to the crop disease recognition model, the training set data is used for training to obtain a trained crop disease recognition model; S4: According to the trained crop disease recognition model, the crop images are processed to complete the crop disease recognition.

2. The crop disease identification method based on adaptive dynamic grouping according to claim 1, characterized in that: The crop disease recognition model comprises an inverse residual module, a first dynamic group recognition module, a second dynamic group recognition module and a classifier connected in sequence; The input training set data of the inverse residual module is fused with the output data of the residual module through a skip connection and used as the input data of the first dynamic group recognition module; The input data of the first dynamic group identification module is merged with the output data of the first dynamic group identification module through a jump connection and used as the input data of the second dynamic group identification module; The input data of the second dynamic group identification module is fused with the output data of the second dynamic group identification module through a skip connection and serves as the input data of the classifier.

3. The crop disease identification method based on adaptive dynamic grouping according to claim 2, characterized in that: The first dynamic group identification module and the second dynamic group identification module have the same structure, and both include a dynamic group convolution submodule, a conditional batch normalization submodule, and a multi-level channel rearrangement submodule connected in sequence; The inverse residual module is used to extract the features of the crop disease image to obtain a feature map; The dynamic group convolution submodule is used to adaptively adjust the number of groups of the group convolution according to the feature map, and perform a convolution operation to obtain a feature map after dynamic group convolution; The conditional batch normalization submodule is used to normalize the feature map after convolution; The multi-level channel rearrangement submodule is used to perform multi-level channel rearrangement on the feature map after dynamic grouping convolution, and output the feature map of multi-level channel rearrangement; The classifier is used to integrate the feature maps of the multi-level channel re-shots input into the classifier and output the crop disease recognition results.

4. The crop disease identification method based on adaptive dynamic grouping according to claim 3 is characterized in that: The inverse residual module includes a 1x1 first convolutional layer, a 3x3 depth-separable convolution, and a 1x1 second convolutional layer connected in sequence; The 1x1 first convolutional layer is used to expand the number of input channels to a high dimension; The 3x3 depthwise separable convolution is used to perform a convolution operation on each channel and extract spatial information; The 1x1 second convolutional layer is used to compress the number of channels back to the dimension of the initial input channels.

5. The crop disease identification method based on adaptive dynamic grouping according to claim 3, characterized in that: The dynamic group convolution submodule adaptively adjusts the number of group convolutions, and the specific expression of the method is as follows: Among them, X is the input feature map, G(X) is the number of groups of group convolution, G min is the minimum number of groups, G max is the maximum number of groups, α is the adjustment coefficient, σ 2 (X) is the global variance of the input feature map, is the upper limit of the input feature variance, is the lower limit of the input feature variance, C in is the number of input channels, is the global variance of the cth channel, H is the height of the input feature map, W is the width of the input feature map, i and j are the indexes of the input feature map, c is the sequence number of the channel, μ c is the mean of the cth channel. The dynamic group convolution submodule performs a dynamic group convolution operation to obtain a feature map after dynamic group convolution, and its specific expression is as follows: Among them, Y g (i,j,c g ) is the feature map after dynamic group convolution, k g (X) is the convolution kernel size of the g-th group convolution, is the number of input channels of the g-th group convolution, m and n are the internal indexes used to traverse the convolution kernel, c ′ is the index of the input channel, i+m-1 and j+n-1 are the translation operations of the convolution kernel on the input feature map, and K g is the convolution kernel configuration for the g-th group convolution, c g is the index of the output channel of the g-th group convolution, k min is the minimum convolution kernel size, k max is the maximum convolution kernel size, σ 2 (X g ) is the variance of the g-th group convolution input features.

6. The crop disease identification method based on adaptive dynamic grouping according to claim 3, characterized in that: The multi-level channel rearrangement is performed on the feature map after the dynamic group convolution. The specific steps of the channel rearrangement are as follows: A1: Perform a reshape operation based on the feature map after dynamic group convolution to obtain a matrix; A2: Perform a transpose operation on the matrix to obtain a transposed matrix; A3: Perform a flattening operation based on the transposed matrix to obtain a feature map with rearranged channels.

7. The crop disease identification method based on adaptive dynamic grouping according to claim 6, characterized in that: The expression of the feature map of the multi-level channel re-shoot is as follows: Y″(i,j,c)=Shuffle(Concat(Y′1(i,j,·),Y′2(i,j,·),…,Y′ G (i,j,·))) =Shuffle(Concat(Shuffle(Y1(i,j,·)),…,Shuffle(Y G (i,j,·)))) Y′ g (i,j,c)=Shuffle(Y g (i,j,c g )) Among them, Y″(i,j,c) is the feature map of multi-level channel re-shooting, Y′ g (i, j, c) is the feature map of the first-level channel rearrangement, Y g (i,j,c g ) is the feature map after dynamic group convolution, Shuffle is the channel reordering, and Concat is the connection operation.

Images