Lightweight large convolution kernel image feature extraction network model and its application

Abstract

The present invention discloses a lightweight large convolution kernel image feature extraction network model and its application, belonging to the field of visual image processing technology. The model comprises a cascade of standard convolution, a first-stage convolution group, a spatial attention module, a second-stage convolution group, a spatial-channel attention module, a third-stage convolution group, a spatial-channel attention module, a fourth-stage convolution group, and a channel attention module. Each stage convolution group comprises a cascade of lightweight multi-branch convolution blocks and an adaptive large-core multi-branch convolution block. The lightweight multi-branch convolution block comprises a pooling layer and several lightweight and efficient hybrid convolution layers. The adaptive large-core multi-branch convolution block comprises an adaptive large-core convolution layer and several lightweight and efficient hybrid convolution layers. The present invention reduces the number of model parameters and computational complexity, lowers computational cost, and maintains the model's real-time inference speed and powerful feature extraction capabilities by simultaneously introducing large-core convolution and an appropriate multi-branch strategy through lightweight convolution operations.

Description

Technical Field

[0001] The invention discloses a lightweight large convolution kernel image feature extraction network, belonging to computer image vision processing technology. Background Art

[0002] Feature extraction is one of the most important issues in computer vision. Its goal is to extract both shallow and deep semantic features from images, enabling subsequent algorithms to run more efficiently and achieve visual tasks such as image defect detection, image classification, and semantic segmentation. Classic deep learning image feature extraction networks can automatically learn features from images and possess powerful feature capture capabilities. However, these networks are bulky, computationally inefficient, slow to infer, and require high computing resources and computational costs.

[0003] Designing deep learning image feature extraction networks using lightweight concepts, such as depthwise separable convolution and grouped convolution, can reduce the computational complexity of the convolution process. Furthermore, replacing fully connected layers with global pooling and implementing channel-wise dimensionality reduction for features using 1×1 convolutions can further reduce the model's computational complexity. However, while these approaches reduce the number of parameters and computational complexity of deep learning image feature extraction network models, improve inference speed, and reduce computing resource requirements, their feature capture capabilities are inferior to those of classic deep learning image feature extraction networks, making them difficult to achieve detailed visual detection or segmentation tasks in complex scenes.

[0004] With the increasing real-time requirements of industrial application scenarios, the acquired image backgrounds are becoming more and more complex, requiring fine defect detection / segmentation and being sensitive to hardware costs. The existing classic image feature extraction network models and lightweight deep learning image feature extraction networks are difficult to meet the above requirements. Summary of the Invention

[0005] The technical problem solved by the present invention is: to provide a lightweight large convolution kernel image feature extraction network to address the problem that the existing image feature extraction network is difficult to simultaneously take into account efficiency, computational cost, feature capture capability, and real-time performance in the feature extraction process of visual detection or segmentation tasks.

[0006] The present invention is implemented by the following technical solutions:

[0007] The present invention first discloses a lightweight large convolution kernel image feature extraction network model, comprising a standard convolution, a first-stage convolution group, a spatial attention module, a second-stage convolution group, a spatial-channel attention module, a third-stage convolution group, a spatial-channel attention module, a fourth-stage convolution group and a channel attention module cascaded in sequence; wherein each stage convolution group comprises a cascaded lightweight multi-branch convolution block LMBlock and an adaptive large-core multi-branch convolution block ALKBlock; the lightweight multi-branch convolution block LMBlock comprises a pooling layer and a plurality of lightweight and efficient hybrid convolution layers LaEConv; the adaptive large-core multi-branch convolution block ALKBlock comprises an adaptive large-core convolution layer and a plurality of lightweight and efficient hybrid convolution layers LaEConv.

[0008] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the lightweight multi-branch convolution block LMBlock performs a channel splitting operation on the first input feature map M, dividing it into three sub-feature maps M1, M2 and M3 with different numbers of channels.

[0009] For the sub-feature map M1, it passes through the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm, SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm and SiLU activation function in sequence to obtain the sub-feature map M1';

[0010] For the sub-feature map M2, through the pooling layer, the sub-feature map M2' is obtained;

[0011] For the sub-feature map M3, the second lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm and the SiLU activation function are sequentially passed to obtain the sub-feature map M3';

[0012] The sub-feature maps M1', M2' and M3' are spliced ​​by channel to obtain the sub-feature map M4. The channel Shuffle operation is performed on the feature map M4 to obtain the first output feature map M out .

[0013] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the lightweight and efficient hybrid convolution layer LaEConv performs a depth-separable convolution DWConv operation on the input feature map x to obtain a feature map x1, performs a partial convolution operation PConv on the feature map x1 to obtain a feature map x2; splices the feature map x1 and the feature map x2 to generate a feature map x3; performs point-by-point multiplication on the feature map x2 and the feature map x3 to obtain a feature map x4; performs a shuffle operation on the feature map x4 to obtain an output feature map x5.

[0014] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the adaptive large kernel multi-branch convolution block ALKBlock performs a channel splitting operation on the feature map output by the lightweight multi-branch convolution block LMBlock, dividing the channel into three sub-feature maps P1, P2 and P3 with different numbers of channels;

[0015] For the sub-feature map P1, it passes through the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the adaptive large kernel convolution layer ALKConv, and the first lightweight and efficient hybrid convolution layer LaEConv in sequence to obtain the sub-feature map P4;

[0016] For the sub-feature map P2, add the sub-feature map P2 and the sub-feature map P4 element by element to obtain the sub-feature map P5;

[0017] For the sub-feature map P3, the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm and SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm and SiLU activation function are sequentially passed to obtain the sub-feature map P6;

[0018] The sub-feature map P4 is activated by the ReLu function to obtain the sub-feature map P4'; the sub-feature maps P5, P6 and P4' are spliced ​​by channel to obtain the sub-feature map P7; the sub-feature map P7 is subjected to the channel Shuffle operation to obtain the second output feature map P out .

[0019] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the convolution kernel size of the first lightweight and efficient hybrid convolution layer LaEConv is 1×1 and the step size is 1; the convolution kernel size of the second lightweight and efficient hybrid convolution layer LaEConv is 3×3 and the step size is 2.

[0020] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the convolution kernel size of the adaptive large kernel convolution layer ALKConv of the adaptive large kernel multi-branch convolution block ALKBlock in the first-stage convolution group is 31×31; the convolution kernel size of the adaptive large kernel convolution layer ALKConv of the adaptive large kernel multi-branch convolution block ALKBlock in the second-stage convolution group is 29×29; the convolution kernel size of the adaptive large kernel convolution layer ALKConv of the adaptive large kernel multi-branch convolution block ALKBlock in the third-stage convolution group is 27×27; the convolution kernel size of the adaptive large kernel convolution layer ALKConv of the adaptive large kernel multi-branch convolution block ALKBlock in the fourth-stage convolution group is 23×23.

[0021] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the pooling layer of the lightweight multi-branch convolution block LMBlock in the first-stage convolution group adopts maximum pooling; the pooling layer of the lightweight multi-branch convolution block LMBlock in the second-stage convolution group adopts maximum pooling; the pooling layer of the lightweight multi-branch convolution block LMBlock in the third-stage convolution group adopts average pooling; in the fourth-stage convolution group, the pooling layer of the lightweight multi-branch convolution block LMBlock adopts average pooling.

[0022] In a lightweight large convolution kernel image feature extraction network model of the present invention, further, the spatial attention module selects the spatial attention module of the convolution block attention mechanism CBAM; the spatial-channel attention module selects the spatial attention module and the channel attention module of the convolution block attention mechanism CBAM in series; the channel attention module selects the channel attention module of the convolution block attention mechanism CBAM.

[0023] The present invention also provides an image feature extraction device, comprising:

[0024] Input module, used to input the original image acquired by the visual sensor,

[0025] A memory module storing a computer program for implementing the aforementioned lightweight large convolution kernel image feature extraction network model of the present invention;

[0026] a processor module, which calls the computer program of the memory module according to the original image input by the input module, and automatically extracts a multi-scale feature map of the original image;

[0027] Output module,outputs the multi-scale feature maps extracted by the processor module.

[0028] The present invention has the following beneficial effects:

[0029] (1) The lightweight multi-branch convolution block LMBlock and the adaptive large-core multi-branch convolution block ALKBlock adopted by the present invention use LaEConv convolution operations, pay attention to the cost of grouped convolution, reduce the degree of fragmentation and reduce element-by-element operations, reduce the number of model parameters and computational complexity, and reduce computational costs. The adaptive large-core multi-branch convolution block ALKBlock introduces a large convolution kernel. Compared with CNNs with small convolution kernels, CNNs with large convolution layers have a larger receptive field and better shape deviation, thereby more effectively capturing features.

[0030] (2) In the lightweight and efficient hybrid convolution layer LaEConv adopted by the present invention, the input data x is first diverted to two data processing paths to generate outputs x1 and x2 respectively. Path one uses depth convolution, focusing on capturing the correlation of local features in the spatial dimension, and each input channel is only connected to one output channel, which significantly reduces the number of parameters. Path two combines depth convolution with point-by-point convolution to form an efficient depth-separable convolution. This design not only further compresses the parameter scale and reduces the computational complexity, but also infiltrates the feature information of depth convolution into the features of point-by-point convolution, thereby enriching the expression of features. Subsequently, x1 and x2 are spliced ​​to generate a new feature map x3. At this time, x2 is cleverly used to perform channel weighting on x3 to achieve adaptive emphasis and suppression of feature channels. Finally, through feature reshaping and shuffle operations, the model can more efficiently integrate and utilize the extracted feature information.

[0031] In summary, the lightweight large convolution kernel image feature extraction network model and its application provided by the present invention construct a lightweight convolution operation LaEConv, while introducing large kernel convolution and appropriate multi-branch strategies, reducing the number of model parameters and computational complexity, reducing computational costs, maintaining real-time inference speed and powerful feature extraction capabilities, and meeting the needs of fine detection and segmentation in complex industrial real-time application scenarios.

[0032] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the overall cascade of the lightweight large convolution kernel image feature extraction network of the present invention.

[0034] Figure 2 Schematic diagram of the cascade of the lightweight multi-branch convolution block LMBlock in the present invention.

[0035] Figure 3 This is a schematic diagram of the cascade of the lightweight and efficient hybrid convolutional layer LaEConv in the present invention.

[0036] Figure 4This is a schematic diagram of the cascade of the adaptive large-kernel multi-branch convolution block ALKBlock in the present invention. DETAILED DESCRIPTION

[0037] Example

[0038] See also Figure 1 The figure shows a specific implementation scheme of the lightweight large convolution kernel image feature extraction network model of the present invention. The image feature extraction network model includes a cascade of standard convolution, a first-stage convolution group, a spatial attention module, a second-stage convolution group, a spatial-channel attention module, a third-stage convolution group, a spatial-channel attention module, a fourth-stage convolution group, and a channel attention module; wherein each stage convolution group includes a cascade of lightweight multi-branch convolution blocks LMBlock and adaptive large-core multi-branch convolution blocks ALKBlock. The image feature extraction network model automatically extracts features based on the input image and ultimately outputs multi-scale feature maps F1, F2, F3, and F4. The feature map F1 is obtained by the spatial attention module, the two spatial-channel attention modules respectively obtain feature maps F2 and F3, and the channel attention module obtains feature map F4.

[0039] Lightweight multi-branch convolution block LMBlock Figure 2 As shown in the figure, the first input feature map M is extracted from the original feature map of the input model by standard convolution. If the number of channels of the first input feature map M is c, the input feature map M is split into three sub-feature maps M1, M2 and M3 with the number of channels c0, c1, and c-c0-c1 respectively. For the sub-feature map M1, the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a step size of 1, the batch normalization layer BatchNorm, the SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a step size of 1, the batch normalization layer BatchNorm and the SiLU activation function are sequentially passed through to obtain the sub-feature map M1'; for the sub-feature map M2, the pooling layer , get the sub-feature map M2'; for the sub-feature map M3, sequentially pass through the second lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a step size of 1, the batch normalization layer BatchNorm and the SiLU activation function to get the sub-feature map M3'; concatenate the sub-feature maps M1', M2' and M3' by channel to get the sub-feature map M4; use the channel Shuffle operation on the feature map M4 to get the first output feature map Mout .

[0040] Lightweight and efficient hybrid convolutional layers such as Figure 3 As shown in the figure, the lightweight and efficient hybrid convolution layer LaEConv takes as input a feature map x. A depthwise separable convolution (DWConv) operation is performed on x to obtain a feature map x1. A partial convolution (PConv) operation is performed on feature map x1 to obtain a feature map x2. Feature maps x1 and x2 are concatenated to generate a feature map x3. Feature maps x2 and x3 are point-wise multiplied to obtain a feature map x4. A shuffle operation is performed on feature map x4 to obtain a feature map x5, which is output by the lightweight and efficient hybrid convolution layer LaEConv. The processes for the first and second lightweight and efficient hybrid convolution layers LaEConv are the same as above.

[0041] Adaptive large kernel multi-branch convolution block ALKBlock Figure 4 As shown in the figure, the input feature map P is the first output feature map M of the lightweight multi-branch convolution block LMBlock out , if the number of channels of the second input feature map P is c, perform a channel splitting operation on the second input feature map P, and divide it into three sub-feature maps P1, P2, and P3 with the number of channels c0, c1, and c-c0-c1 respectively. For the sub-feature map P1, pass through the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a stride of 1, the adaptive large kernel convolution layer ALKConv with a convolution kernel size of K×K, and the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a stride of 1 to obtain the sub-feature map P4; add the sub-feature map P2 and the sub-feature map P4 element by element to obtain the sub-feature map P5; for the sub-feature map P3, pass through the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a stride of 1, the batch normalization layer BatchN orm and SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv with a convolution kernel size of 1×1 and a step size of 1, the batch normalization layer BatchNorm and SiLU activation function, to obtain the sub-feature map P6; the sub-feature map P4 is activated by the ReLu function to obtain the sub-feature map P4'; the sub-feature maps P5, P6 and P4' are spliced ​​by channel to obtain the sub-feature map P7; the sub-feature map P7 is subjected to the channel Shuffle operation to obtain the second output feature map P out .

[0042] In this embodiment, the convolution kernel size of the adaptive large-core convolution layer ALKConv of the adaptive large-core multi-branch convolution block ALKBlock in the first-stage convolution group is 31×31; the convolution kernel size of the adaptive large-core convolution layer ALKConv of the adaptive large-core multi-branch convolution block ALKBlock in the second-stage convolution group is 29×29; the convolution kernel size of the adaptive large-core convolution layer ALKConv of the adaptive large-core multi-branch convolution block ALKBlock in the third-stage convolution group is 27×27; and the convolution kernel size of the adaptive large-core convolution layer ALKConv of the adaptive large-core multi-branch convolution block ALKBlock in the fourth-stage convolution group is 23×23. The stride size of the adaptive large-core convolution layer ALKConv in each stage is 1.

[0043] The pooling layer of the lightweight multi-branch convolution block LMBlock in the first-stage convolution group adopts maximum pooling; the pooling layer of the lightweight multi-branch convolution block LMBlock in the second-stage convolution group adopts maximum pooling; the pooling layer of the lightweight multi-branch convolution block LMBlock in the third-stage convolution group adopts average pooling; in the fourth-stage convolution group, the pooling layer of the lightweight multi-branch convolution block LMBlock adopts average pooling.

[0044] The spatial attention module selects the spatial attention module of the convolutional block attention mechanism (CBAM); the spatial-channel attention selects the spatial attention module and the channel attention module of the convolutional block attention mechanism (CBAM) in series; the channel attention module selects the channel attention module of the convolutional block attention mechanism (CBAM).

[0045] The present invention selects the feature extraction backbone network in the current mainstream detection model for comparative experiments. The compared networks include CFPNet, ELAN, GELAN, ShuffleNet, CSPNeXt, PicoNet, MobileOne, MSYOLO, Shift-ConvNets, and DCNetv2. The visual inspection task for verification is to detect substation equipment from infrared images obtained from drones. There are a total of 1,861 infrared images of substation equipment, including seven types of substation equipment: lightning arrester 1, lightning arrester 2, current transformer 1, current transformer 2, voltage transformer, disconnector, and pillar porcelain bottle. Taking YOLOv5 as the benchmark, the image feature extraction network in the above model is replaced by the image feature extraction network of YOLOv5, and the lightweight large convolution kernel image feature extraction network proposed in the present invention is replaced by the image feature extraction network of YOLOv5 (denoted as LaBNet) to detect the substation equipment in the above dataset. The results are shown in Table 1.

[0046] Table 1 Comparison of the LaBNet of the present invention with other mainstream image feature extraction networks

[0047]

[0048]

[0049] From the data in Table 1, it can be seen that the lightweight large convolution kernel image feature extraction network proposed in this paper has a high mAP 50 It achieved the best performance in terms of metrics, reaching 96.3%. It also ranked third in terms of parameter count, at only 4.92M, with minimal computational complexity and the fastest inference speed.

[0050] In some embodiments, an image feature extraction device is further provided, comprising:

[0051] Input module, used to input the original image acquired by the visual sensor,

[0052] A memory module storing a computer program for implementing a lightweight large convolution kernel image feature extraction network model in this embodiment;

[0053] a processor module, which calls the computer program of the memory module according to the original image input by the input module, and automatically extracts a multi-scale feature map of the original image;

[0054] Output module,outputs the multi-scale feature maps extracted by the processor module.

[0055] The memory module is a computer-readable storage medium, which can be an internal storage unit of the software and hardware device described in any of the aforementioned embodiments, such as a hard disk or memory of a controller. The readable storage medium can also be an external storage device of the controller, such as a plug-in hard disk equipped on the controller, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Furthermore, the readable storage medium can also include both an internal storage unit of the controller and an external storage device. The readable storage medium is used to store the computer program and other programs and data required by the controller. The readable storage medium can also be used to temporarily store data that has been output or is to be output.

[0056] Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned memory module also includes various media that can store program code, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0057] In this document, the directions or positional relationships indicated by terms such as "up", "down", "front", "back", "left", "right", "top", "bottom", "inside", "outside", "vertical", and "horizontal" are based on the directions or positional relationships shown in the accompanying drawings and are only for the clarity of the technical solution and the convenience of description, and therefore should not be understood as limiting the present invention.

[0058] As used herein, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion of elements other than the listed elements and may also include additional elements not specifically listed.

[0059] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A lightweight large convolution kernel image feature extraction network model that extracts multi-scale feature maps of the original image from the original image acquired by the visual sensor, characterized by: The model includes a cascade of standard convolution, a first-stage convolution group, a spatial attention module, a second-stage convolution group, a spatial-channel attention module, a third-stage convolution group, a spatial-channel attention module, a fourth-stage convolution group, and a channel attention module; The convolution groups at each stage include cascaded lightweight multi-branch convolution blocks LMBlock and adaptive large-core multi-branch convolution blocks ALKBlock; The lightweight multi-branch convolution block LMBlock includes a pooling layer and several lightweight and efficient hybrid convolution layers LaEConv, performs a channel splitting operation on the first input feature map M, and divides it into three sub-feature maps M1, M2 and M3 with different numbers of channels from the channel; for the sub-feature map M1, it passes through the first lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm, the SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer Batch Norm and SiLU activation functions are used to obtain the sub-feature map M1'; for the sub-feature map M2, the sub-feature map M2' is obtained through the pooling layer; for the sub-feature map M3, the second lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the batch normalization layer BatchNorm and the SiLU activation function are sequentially passed to obtain the sub-feature map M3'; the sub-feature maps M1', M2' and M3' are spliced ​​by channel to obtain the sub-feature map M4, and the channel Shuffle operation is performed on the feature map M4 to obtain the first output feature map M out ; The adaptive large-kernel multi-branch convolution block ALKBlock includes an adaptive large-kernel convolution layer and several lightweight and efficient hybrid convolution layers LaEConv. It performs a channel splitting operation on the feature map output by the lightweight multi-branch convolution block LMBlock, and divides the feature map into three sub-feature maps P1, P2 and P3 with different numbers of channels. For the sub-feature map P1, the batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, the adaptive large-kernel convolution layer ALKConv, and the first lightweight and efficient hybrid convolution layer LaEConv are sequentially passed through to obtain the sub-feature map P4. For the sub-feature map P2, the sub-feature map P2 and the sub-feature map P4 are added element by element to obtain the sub-feature map P5. ; For the sub-feature map P3, the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm and SiLU activation function, the second lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm, the first lightweight and efficient hybrid convolution layer LaEConv, batch normalization layer BatchNorm and SiLU activation function are sequentially passed to obtain the sub-feature map P6; the sub-feature map P4 is passed through the ReLu activation function to obtain the sub-feature map P4'; the sub-feature maps P5, P6 and P4' are spliced ​​by channel to obtain the sub-feature map P7; the sub-feature map P7 is subjected to the channel Shuffle operation to obtain the second output feature map P out .

2. The lightweight large convolution kernel image feature extraction network model according to claim 1, characterized in that: The lightweight and efficient hybrid convolution layer LaEConv performs a depth-wise separable convolution DWConv operation on the input feature map x to obtain a feature map x1, performs a partial convolution operation PConv on the feature map x1 to obtain a feature map x2; concatenates the feature map x1 and the feature map x2 to generate a feature map x3; performs point-by-point multiplication on the feature map x2 and the feature map x3 to obtain a feature map x4; performs a shuffle operation on the feature map x4 to obtain an output feature map x5.

3. The lightweight large convolution kernel image feature extraction network model according to claim 1, characterized in that: The convolution kernel size of the first lightweight and efficient hybrid convolution layer LaEConv is 1×1 and the step size is 1; the convolution kernel size of the second lightweight and efficient hybrid convolution layer LaEConv is 3×3 and the step size is 2.

4. The lightweight large convolution kernel image feature extraction network model according to claim 1, characterized in that: The convolution kernel size of the adaptive large-kernel convolution layer ALKConv of the adaptive large-kernel multi-branch convolution block ALKBlock in the first-stage convolution group is 31×31; The convolution kernel size of the adaptive large kernel convolution layer ALKConv of the adaptive large kernel multi-branch convolution block ALKBlock in the second stage convolution group is 29×29; The convolution kernel size of the adaptive large-kernel convolution layer ALKConv of the adaptive large-kernel multi-branch convolution block ALKBlock in the third-stage convolution group is 27×27; The convolution kernel size of the adaptive large-kernel convolution layer ALKConv of the adaptive large-kernel multi-branch convolution block ALKBlock in the fourth-stage convolution group is 23×23.

5. The lightweight large convolution kernel image feature extraction network model according to claim 1, characterized in that: The pooling layer of the lightweight multi-branch convolution block LMBlock in the first stage convolution group adopts maximum pooling; The pooling layer of the lightweight multi-branch convolution block LMBlock in the second-stage convolution group adopts maximum pooling; The pooling layer of the lightweight multi-branch convolution block LMBlock in the third stage convolution group adopts average pooling; In the fourth stage convolution group, the pooling layer of the lightweight multi-branch convolution block LMBlock adopts average pooling.

6. The lightweight large convolution kernel image feature extraction network model according to claim 1, characterized in that: The spatial attention module uses the convolutional block attention mechanism CBAM’s spatial attention module; The spatial-channel attention module uses the convolutional block attention mechanism CBAM's spatial attention module and channel attention module in series; The channel attention module uses the channel attention module of the convolutional block attention mechanism CBAM.

7. An image feature extraction device, characterized in that: include: Input module, used to input the original image acquired by the visual sensor, A memory module storing a computer program for implementing a lightweight large convolution kernel image feature extraction network model according to any one of claims 1 to 6; a processor module, which calls the computer program of the memory module according to the original image input by the input module, and automatically extracts a multi-scale feature map of the original image; Output module,outputs the multi-scale feature maps extracted by the processor module.

Images