Image recognition method based on dense connection attention mechanism
By introducing a densely connected attention mechanism into the image recognition method, the problem of over-learning of attention mechanism and insufficient extraction of sub-important features is solved, the image classification accuracy is improved, the model complexity is reduced, and more efficient recognition effect is achieved.
Patent Information
- Application Number
- CN202411040595.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-31
- Publication Date
- 2024-11-19
AI Technical Summary
The existing image classification recognition method based on deep learning has problems such as over-learning of attention mechanism, insufficient extraction of sub-important features, insufficient extraction of attention relationship features within the model, and increasing the complexity of the model in relational networks.
The image recognition method based on the densely connected attention mechanism is adopted, and the convolutional module, attention feature fusion module and densely connected attention relationship convolution group are constructed to realize the fusion and reconstruction of attention features and reduce the complexity of the model.
The image classification accuracy is improved, attention characteristics are evenly distributed, the recognition effect of the model is improved, and the model parameters and calculation amount added in actual applications are less than 0.001%, which has little impact on the recognition efficiency.
Smart Images

Figure CN118982689A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of image classification and recognition, and more specifically, relates to an image recognition method based on a densely connected attention mechanism. Background Art
[0002] Vision plays a vital role in the human cognition of the world. The human eye is not just a static image capture device, but a dynamic, active information detector. When humans observe a scene, the eyeballs constantly make tiny jump movements - called saccades, which allows our vision to quickly scan and focus on the most informative parts of the environment in a limited time. This selective attention not only saves brain processing resources, but also greatly improves our understanding and response speed to complex scenes. For example, when we are on a busy city street, our eyes automatically lock on key elements such as pedestrians, vehicles and traffic signals, while ignoring non-important details in the background, such as the texture of the building or the color of the sky. This process shows how the human visual system efficiently filters out the most relevant information in the environment to form an accurate understanding of the entire scene.
[0003] Inspired by the human visual system, researchers began to explore how to implement similar selective attention mechanisms in machine learning models. In the field of computer vision, especially in deep learning models, convolutional neural networks (CNNs) have become mainstream technology due to their powerful image feature extraction capabilities. However, traditional CNNs often treat all input pixels as equally important, which makes the model inefficient when processing complex images containing a lot of irrelevant information. To solve this problem, researchers began to introduce attention mechanisms into the CNN architecture. The attention mechanism allows the network to dynamically assign different weights to different areas when processing images, thereby simulating the focus shift of human vision, effectively suppressing irrelevant background noise, and highlighting key features in the image. This mechanism significantly improves the performance of CNN in tasks such as image classification, object detection, and semantic segmentation, because it enables the network to more accurately locate and analyze important parts of the image, thereby making more accurate decisions. However, the attention mechanism has the problem of over-learning during model training, resulting in insufficient learning of sub-important features that affect the classification and recognition results of the model's attention, and designing an attention mechanism relationship network will increase model parameters and reduce recognition efficiency.
[0004] After searching domestic and foreign literature and patents, it was found that the existing patent "Image Target Detection Method Based on Dual Attention Mechanism" (application number: CN202210614509.X) discloses an image target detection method based on a dual attention mechanism. The model obtains a multi-scale feature map by constructing a bottom-up pyramid structure, and embeds the multi-scale feature map into the detection module network of the target detection model based on the dual attention mechanism to obtain a dual attention information feature map including spatial attention information and channel attention information. The candidate region of the input sample is generated according to the dual attention information feature map. The dual attention mechanism uses a multi-directional feature fusion mechanism to send multi-directional fusion features containing rich contextual information into the detection network, and deploys the attention mechanism in the channel and spatial dimensions, which improves the network's ability to focus on useful information and generalization ability, but this also suppresses the less important information related to recognition detection during the model learning process, increases the risk of overfitting of the model, and reduces the recognition detection effect of the model.
[0005] The existing patent "An autonomous driving image recognition method based on attention mechanism and relationship network" (application number: CN202110043727.8) discloses an autonomous driving image recognition method based on attention mechanism and relationship network. This method first uses a convolutional neural network to extract autonomous driving image features, then calculates attention features through the attention mechanism, and then uses the relationship network to generate the relationship results between objects in the image. Finally, the network is trained through the loss function, and the trained model is used to make relationship judgments in autonomous driving scenes. However, this method requires an additional designed relationship network to learn the important features of the image extracted by the attention mechanism, which increases the complexity of the model and reduces the efficiency of model learning.
[0006] In the existing paper "Image Classification Algorithm Based on Split Channel Attention Network" (Journal of Engineering Science, 2024), a split channel attention mechanism is proposed to build a module, which expands the output dimension of global average pooling, reduces the information loss caused by global average pooling, and enhances the feature diversity of the global average pooling layer in channel attention. Then, multiple one-dimensional convolutions are used to calculate the attention weights of each region in the channel dimension. The split channel attention mechanism is combined with a variety of image classification networks. Experimental results on commonly used image classification datasets show that the split channel attention mechanism can still effectively improve the accuracy of the model while maintaining lightweight, and also shows better advantages compared with other attention mechanisms. This model improves image classification performance by improving the channel attention mechanism, but it is insufficient in extracting less important features of the input image, which is not conducive to classification scenarios that require more information.
[0007] In the existing paper "Halftone Image Classification Based on Deep Network with Fusion Attention Mechanism" (Journal of the Chinese Society of Agricultural Machinery, Vol. 42, No. 3, 2024), the channel attention mechanism is used to extract the noise distribution characteristics of halftone images between different channels, and the spatial attention mechanism is used to extract the relationship between the noise points of halftone images in different channels and spaces to classify the identified halftone image noise distribution characteristics, thereby realizing the classification of halftone images. Experimental results show that this method can achieve better classification results than other similar methods. Although the model improves the classification effect of the model by embedding the channel attention mechanism and the spatial attention mechanism, there is still a problem of insufficient consideration of the relationship between the internal attention of the model.
[0008] In summary, the existing image classification and recognition methods based on deep learning attention mechanism have the following shortcomings:
[0009] First, the existing method has the problem of over-learning in the attention mechanism, which leads to insufficient extraction of sub-important features that affect the image classification results;
[0010] Second, existing methods are insufficient in extracting relational features of the attention mechanism within the model;
[0011] 3. The attention mechanism relationship network designed in addition increases the complexity of the model and reduces the learning efficiency.
[0012] Therefore, a new image recognition method based on densely connected attention mechanism is urgently needed. Summary of the invention
[0013] In order to solve the above problems, the present invention proposes an image recognition method based on densely connected attention mechanism.
[0014] In order to solve at least one of the above technical problems, according to one aspect of the present invention, an image recognition method based on a densely connected attention mechanism is provided, comprising the following steps:
[0015] S1, build a convolution module to extract input features;
[0016] S2, build attention feature fusion module;
[0017] S3, superimpose the attention feature fusion module in S2, and densely connect the attention mechanism part to construct a densely connected attention relationship convolution group;
[0018] S4, by stacking densely connected attention relationship convolution groups, a recognition model is formed.
[0019] Furthermore, the specific steps of S1 are:
[0020] The original input image passes through a convolution layer and a pooling window to obtain the feature map of the original input image; the feature map of the original input image is input into the first densely connected attention relationship convolution group; each densely connected attention relationship convolution group contains multiple convolution blocks, and the input feature map passes through each convolution block in turn; the input feature map obtains the output feature map after passing through each convolution block, and the output feature map will be used as the input of the attention feature fusion module of S2.
[0021] Furthermore, in S2,
[0022] The attention feature fusion module includes: feature map compression layer, aggregation convolution operation layer and feature reconstruction layer. The input features pass through each layer in turn.
[0023] Furthermore, the feature map compression layer is composed of a global average pooling layer;
[0024] The aggregate convolution operation layer is composed of a convolution layer including a convolution kernel size of 1×1, a normalization layer and a ReLu activation layer;
[0025] The feature reconstruction layer is composed of a first convolution layer including a convolution kernel size of 1×1, a ReLu activation layer and a second convolution layer including a convolution kernel size of 1×1, which are sequentially connected.
[0026] Furthermore, the specific steps of S2 are:
[0027] S21, the input feature map size in the feature map compression layer has three dimensions. After the global average pooling layer, the size of C×H×W is converted to C×1×1. Then, the channel dimension and the last dimension of the feature with the size of C×1×1 are exchanged to obtain a size of 1×1×C.
[0028] Where C represents the number of channels in the channel dimension, H represents the height of the feature map of the current module input, W represents the width of the feature map of the current module input, and M represents the number of features of different scales;
[0029] S22. No feature aggregation operation is required in the first residual block in each densely connected attention relationship convolution group; starting from the second residual block, the global feature vector of size 1×1×C after compression and conversion by the global average pooling layer is merged with the feature vector of size 1×1×C output by the feature compression layer of the current residual module in the channel dimension, that is, a feature vector of size 2×1×C is obtained; and by analogy, the subsequent residual blocks will aggregate all the feature vectors after feature compression, thereby obtaining a feature vector of size n×1×C, where n represents the sequence number of the current residual block in the residual group;
[0030] S23, the number of input channels of the aggregate convolution operation layer is n+1, and the number of output channels is 1. After passing through the convolution layer, normalization layer, and activation layer in sequence, all the previous features can be associated and fused, and the channel dimension is compressed to 1, and the fused attention feature with an output size of 1×1×C is obtained. Then, the dimensions are exchanged to obtain a feature vector with a size of C×1×1.
[0031] S24: The output of the aggregated convolution operation layer is input into a convolution layer with a convolution kernel size of 1×1 and dimensionally compressed in the channel dimension so that the output feature map size becomes C / r×H×W.
[0032] Where r represents the compression ratio;
[0033] S25: The output feature map is input into another convolutional layer with a convolution kernel size of 1×1 to restore the number of feature map dimensions to the original size, that is, a feature vector of size C×H×W. Then, the corresponding position is multiplied by the output feature vector of S1 through matrix operation to reconstruct the output features.
[0034] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the steps of the image recognition method based on the densely connected attention mechanism of the present invention are implemented.
[0035] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the image recognition method based on the densely connected attention mechanism of the present invention are implemented.
[0036] Compared with the prior art, the above method of the present invention has the following beneficial effects:
[0037] (1) The method of the present invention can be applied to different types of attention mechanisms to improve the effect of the attention mechanism, and has achieved better classification accuracy than the original attention mechanism method on the extracted data sets.
[0038] (2) The method proposed in the present invention can connect and fuse the attention features between layers in the model, so that the model can distribute the attention more evenly during the learning process, thereby obtaining more secondary important features and improving the classification and recognition effect of the model.
[0039] (3) The method of the present invention is used to connect the attention modules on the basis of the original attention model. The impact of the increased model parameters and calculation amount is very small, about 0.001%, which can be ignored in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments will be briefly introduced below. Obviously, the drawings in the following description only relate to some embodiments of the present invention, but are not intended to limit the present invention.
[0041] Figure 1 A schematic diagram of a densely connected attention relationship network structure of a preferred embodiment of the present invention;
[0042] Figure 2 A schematic diagram of a densely connected attention relationship convolution group including 5 residual blocks in a preferred embodiment of the present invention;
[0043] Figure 3 A flowchart of an image recognition method based on a densely connected attention mechanism in a preferred embodiment of the present invention;
[0044] Figure 4 The figure is a schematic diagram of experimental results of a preferred embodiment of the present invention. DETAILED DESCRIPTION
[0045] To make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the embodiments of the present invention will be described clearly and completely in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.
[0046] Unless otherwise defined, technical or scientific terms used herein shall have the same meaning as commonly understood by one of ordinary skill in the art to which the invention belongs.
[0047] Embodiment 1:
[0048] like Figure 1-4 As shown, the present invention proposes an image recognition method based on a densely connected attention mechanism, comprising the following steps:
[0049] S1, build a convolution module to extract input features;
[0050] S2, build attention feature fusion module;
[0051] S3, superimpose the attention feature fusion module in S2, and densely connect the attention mechanism part to construct a densely connected attention relationship convolution group;
[0052] S4, by stacking densely connected attention relationship convolution groups, a recognition model is formed;
[0053] S1 specifically includes:
[0054] This method is based on the existing convolutional neural network model and is illustrated using the Squeezed and Excited Residual Network (SE-ResNet) model as an example.
[0055] The original input image size is 224×224 pixels. After passing through a convolution layer with a convolution kernel size of 7×7 pixels and a stride of 2 (i.e., each time the convolution operation is performed by shifting two pixels), the input image with 3 channels is converted into a feature map with 64 channels and a size of 112×112 pixels. After passing through a maximum pooling layer with a pooling window of 3×3 pixels, a stride of 2, and a padding of 1, the output size becomes 56×56 pixels, and the feature map is input into the first densely connected attention relationship convolution group. Each densely connected attention relationship convolution group contains multiple convolution blocks, and the input feature map passes through each convolution block in turn; after the input feature map passes through each convolution block, the output feature map is obtained, and the output feature map will be used as the input of the attention feature fusion module of S2.
[0056] There are four residual groups in the SE-ResNet model. The output feature map sizes of the second, third and fourth groups are 28×28 pixels, 14×14 pixels and 7×7 pixels, and the number of channels are 512, 1024 and 2048, respectively.
[0057] S2 specifically includes:
[0058] In the attention feature fusion module, there are four parts: feature map compression and aggregation layer, aggregation convolution operation layer and feature reconstruction layer. The input features will pass through each layer in turn.
[0059] S21: The feature map compression layer consists of a global average pooling layer. The input feature map size is three dimensions in total. After the global average pooling layer, the size is converted from C×H×W to G×1×1. Then the channel dimension and the last dimension of the feature are exchanged to obtain a size of 1×1×C.
[0060] Where C represents the number of channels in the channel dimension, H represents the height of the feature map of the current module input, W represents the width of the feature map of the current module input, and M represents the number of features of different scales;
[0061] S22: No feature aggregation operation is required in the first residual block in each densely connected attention relationship convolution group. Starting from the second residual block, the global feature vector of size 1×1×C after compression and conversion by the global average pooling layer is merged with the feature vector of size 1×1×C output by the feature compression layer of the current residual module in the channel dimension, that is, a feature vector of size 2×1×C is obtained. By analogy, the subsequent residual blocks will aggregate all the feature vectors after feature compression, thereby obtaining a feature vector of size n×1×C, where n represents the sequence number of the current residual block in the residual group.
[0062] The aggregate convolution operation layer consists of a convolution layer with a convolution kernel size of 1×1, a normalization layer, and a ReLu activation layer.
[0063] S23: The number of input channels of the aggregate convolution operation layer is n+1, and the number of output channels is 1. After passing through the convolution layer, normalization layer, and activation layer in sequence, all the previous features can be associated and fused, and the channel dimension is compressed to 1, so that the output size of the fused attention feature is 1×1×C, and then the dimension is exchanged to obtain a feature vector of size C×1×1;
[0064] The feature reconstruction layer consists of a first convolutional layer with a convolution kernel size of 1×1, a ReLu activation layer, and a second convolutional layer with a convolution kernel size of 1×1, connected in sequence.
[0065] S24: In order to further integrate the attention features of each layer and enhance the expression performance of the attention features, the output of the aggregated convolution operation layer is input into a convolution layer with a convolution kernel size of 1×1 and dimensionally compressed in the channel dimension so that the output feature map size becomes C / r×H×W;
[0066] Where r represents the compression ratio;
[0067] S25: Input the output feature map into another convolutional layer with a convolution kernel size of 1×1 to restore the number of feature map dimensions to the original size, that is, a feature vector of size C×H×W, and then multiply the corresponding position by the output feature vector of step 1 through matrix operation to reconstruct its features;
[0068] To verify the effect of the model of the present invention, some pictures from the large public image dataset ImageNet2012 were extracted for verification. 500 classes were randomly selected from the training set, and half of the images were extracted from each class. The verification set extracted all the pictures of the corresponding 500 classes as the model verification set. The initial value of the network model was randomly set during model training, and the average recognition accuracy of the results of 5 training tests was compared. The experimental results are shown in Figure 2. Figure 4 As shown, by comparison, it can be found that the accuracy performance that can be obtained by the method of the present invention is more stable than that of other methods, thereby proving the robustness of the method.
[0069] For the residual network model using the SE attention mechanism method, the Top1 accuracy of the model can be improved by 2.21% after the method of the present invention is adopted. For the residual network model using the ECA attention mechanism method, the Top1 accuracy of the model can be improved by 2.47% after the method of the present invention is adopted. These experimental results prove the effectiveness of this method. The experimental results are shown in Table 1. Among them, ResNet+SE means that the SE attention mechanism is embedded in the ResNet model, and ResNet+SE+DCA means that the method of the present invention is added on this basis. ResNet+ECA means that the ECA attention mechanism is embedded in the ResNet model, and ResNet+ECA+DCA means that the method of the present invention is added on this basis.
[0070] Table 1 is the average accuracy comparison results of the image recognition method based on the densely connected attention mechanism of the present invention:
[0071]
[0072] Embodiment 2:
[0073] The computer-readable storage medium of this embodiment stores a computer program thereon, which, when executed by a processor, implements the steps in the image recognition method based on the densely connected attention mechanism of Example 1.
[0074] The computer-readable storage medium of this embodiment may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal; the computer-readable storage medium of this embodiment may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card, a secure digital card, a flash memory card, etc. equipped on the terminal; further, the computer-readable storage medium may also include both an internal storage unit of the terminal and an external storage device.
[0075] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or is to be output.
[0076] Embodiment 3:
[0077] The computer device of this embodiment includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the image recognition method based on the densely connected attention mechanism of Embodiment 1 are implemented.
[0078] In this embodiment, the processor may be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, readily available programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.
[0079] Those skilled in the art will appreciate that the disclosed content of the embodiments may be provided as methods, systems, or computer program products. Therefore, the present solution may adopt the form of hardware embodiments, software embodiments, or embodiments combining software and hardware. Moreover, the present solution may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) containing computer-usable program codes.
[0080] The present solution is described with reference to the flowchart and / or block diagram of the method and computer program product according to the embodiment of the present solution. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions; these computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0081] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0082] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0083] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.
[0084] The examples described in the present invention are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should all fall within the protection scope of the present invention.
Claims
1. An image recognition method based on densely connected attention mechanism, characterized in that: The steps include: S1, build a convolution module to extract input features; S2, build attention feature fusion module; S3, superimpose the attention feature fusion module in S2, and densely connect the attention mechanism part to construct a densely connected attention relationship convolution group; S4, by stacking densely connected attention relationship convolution groups, a recognition model is formed.
2. The method according to claim 1, characterized in that The specific steps of S1 are: The original input image passes through a convolution layer and a pooling window to obtain the feature map of the original input image; the feature map of the original input image is input into the first densely connected attention relationship convolution group; each densely connected attention relationship convolution group contains multiple convolution blocks, and the input feature map passes through each convolution block in turn; the input feature map obtains the output feature map after passing through each convolution block, and the output feature map will be used as the input of the attention feature fusion module of S2.
3. The method according to claim 2, characterized in that In S2, The attention feature fusion module includes: feature map compression layer, aggregation convolution operation layer and feature reconstruction layer. The input features pass through each layer in turn.
4. The method according to claim 3, characterized in that: The feature map compression layer is composed of a global average pooling layer; The aggregate convolution operation layer is composed of a convolution layer including a convolution kernel size of 1×1, a normalization layer and a ReLu activation layer; The feature reconstruction layer is composed of a first convolution layer including a convolution kernel size of 1×1, a ReLu activation layer and a second convolution layer including a convolution kernel size of 1×1, which are sequentially connected.
5. The method according to claim 4, characterized in that The specific steps of S2 are: S21, the input feature map size in the feature map compression layer has three dimensions. After the global average pooling layer, the size of C×H×W is converted to C×1×1. Then, the channel dimension and the last dimension of the feature with the size of C×1×1 are exchanged to obtain a size of 1×1×C. Where C represents the number of channels in the channel dimension, H represents the height of the feature map of the current module input, W represents the width of the feature map of the current module input, and M represents the number of features of different scales; S22. No feature aggregation operation is required in the first residual block in each densely connected attention relationship convolution group; starting from the second residual block, the global feature vector of size 1×1×C after compression and conversion by the global average pooling layer is merged with the feature vector of size 1×1×C output by the feature compression layer of the current residual module in the channel dimension, that is, a feature vector of size 2×1×C is obtained; and by analogy, the subsequent residual blocks will aggregate all the feature vectors after feature compression, thereby obtaining a feature vector of size n×1×C, where n represents the sequence number of the current residual block in the residual group; S23, the number of input channels of the aggregate convolution operation layer is n+1, and the number of output channels is 1. After passing through the convolution layer, normalization layer, and activation layer in sequence, all the previous features can be associated and fused, and the channel dimension is compressed to 1, and the fused attention feature with an output size of 1×1×C is obtained. Then, the dimensions are exchanged to obtain a feature vector with a size of C×1×1. S24: The output of the aggregated convolution operation layer is input into a convolution layer with a convolution kernel size of 1×1 and dimensionally compressed in the channel dimension so that the output feature map size becomes C / r×H×W. Where r represents the compression ratio; S25: The output feature map is input into another convolutional layer with a convolution kernel size of 1×1 to restore the number of feature map dimensions to the original size, that is, a feature vector of size C×H×W. Then, the corresponding position is multiplied by the output feature vector of S1 through matrix operation to reconstruct the output features.
6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps in the image recognition method based on the densely connected attention mechanism as described in any one of claims 1 to 5 are implemented.
7. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the image recognition method based on the densely connected attention mechanism as described in any one of claims 1 to 5 are implemented.
