Lightweight self-adaptive collaborative attention mechanism

Abstract

A lightweight self-adaptive collaborative attention mechanism is characterized in that input features are divided into n different groups, each group is independently processed through global average pooling, self-adaptive convolution and a sigmoid activation function to obtain the attention weight of the group, then the attention weight is multiplied by the features of the group to obtain the weighted features of the group, and the weighted features of the group are calculated. And finally, performing concat splicing on each group of weighted features to generate final output. The lightweight adaptive collaborative attention mechanism has the beneficial effects that the lightweight adaptive collaborative attention mechanism can enhance the adaptability of the model, not only can reduce redundant calculation, but also can perform multi-scale feature enhancement.

Description

Technical Field

[0001] The present invention relates to the field of ACAM, and in particular to a lightweight adaptive collaborative attention mechanism. Background Art

[0002] As the latest generation of YOLO series algorithms, YOLOv10n has high detection accuracy and fast reasoning speed, and is suitable for various target detection tasks. However, for transmission line insulator defect detection, YOLOv10n still has some shortcomings: see Figure 2 As shown in the figure, for similar insulator flashover discharge defect types, the defect target is small and the defect features are not obvious. However, the original model's early feature extraction capability is limited, resulting in low detection accuracy or even missed detection when the model detects such targets. Summary of the invention

[0003] The purpose of the present invention is to provide a lightweight adaptive collaborative attention mechanism to address the deficiencies and defects in the prior art. The lightweight adaptive collaborative attention mechanism has the ability to enhance the adaptability of the model, which can not only reduce redundant calculations, but also enhance multi-scale features.

[0004] To achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a lightweight adaptive collaborative attention mechanism, which specifically comprises: dividing the input features into n different groups, each group is independently processed by global average pooling, adaptive convolution and sigmoid activation function to obtain the attention weight of the group, and then the attention weight is multiplied by the group of features to obtain the weighted features of the group, and finally the weighted groups of features are concat spliced ​​to generate the final output.

[0005] Furthermore, the size of the adaptive convolution kernel refers to the setting of the one-dimensional convolution kernel size in the ECA attention mechanism, and the average value of the variance of each set of feature channels is introduced on this basis. Let the convolution kernel size be dynamically adjusted according to the variance change of the input feature map.

[0006] Furthermore, it includes the following calculation formula: Among them, X i is the input feature map of the i-th group, n is the number of feature groups; f k×k is a convolution operation with a convolution kernel size of k; is the average value of all channel variances; G is the number of channels in each group; γ is the adjustment factor.

[0007] After adopting the above technical solution, the beneficial effects of the present invention are: the lightweight adaptive collaborative attention mechanism has at least the following advantages:

[0008] 1. Enhance the adaptability of the model: By dynamically adjusting the convolution kernel size according to the variance, the model can adaptively capture subtle or significant changes in the feature map. This adaptability enables the model to flexibly adjust the receptive field of its convolution operation when processing different types of inputs (such as images with different texture complexity), thereby better capturing important features.

[0009] 2. Reduce redundant calculations: If the variance of the input feature map is small, it means that the feature set has a small change and is a secondary feature. In this case, a smaller convolution kernel can be selected to reduce the amount of calculation. Conversely, when the variance is large, the model may need to use a larger convolution kernel to capture more contextual information. This dynamic adjustment can effectively avoid excessive or insufficient calculations and improve overall computing efficiency.

[0010] 3. Multi-scale feature enhancement: Different convolution kernel sizes can capture features of different scales. By dynamically adjusting the convolution kernel size, the model can enhance features of different scales separately. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0012] Figure 1 It is a structural schematic diagram of the present invention.

[0013] Figure 2 This is a detection problem existing in the prior art.

[0014] Figure 3 This is a heat map visualization comparison of the initial YOLOv10n network in the present invention.

[0015] Figure 4 It is a visualization comparison diagram of the heat map of the YOLOv10+ACAM network in the present invention.

[0016] Figure 5 This is the initial YOLOv10n detection result diagram in the present invention.

[0017] Figure 6 This is the YOLOv10n+ACAM detection result diagram in the present invention. DETAILED DESCRIPTION

[0018] See also Figure 1As shown, the technical solution adopted in this specific implementation is: it specifically comprises: dividing the input features into n different groups, each group is independently processed by global average pooling, adaptive convolution and sigmoid activation function to obtain the attention weight of the group, and then the attention weight is multiplied by the group of features to obtain the weighted features of the group, and finally the weighted groups of features are concat spliced ​​to generate the final output, wherein the size of the adaptive convolution kernel refers to the setting of the one-dimensional convolution kernel size in the ECA attention mechanism, and the average value of the variance of each group of feature channels is introduced on this basis Let the convolution kernel size be dynamically adjusted according to the variance change of the input feature map.

[0019] It includes the following calculation formula:

[0020]

[0021] Among them, X i is the input feature map of the i-th group, n is the number of feature groups; f k×k is a convolution operation with a convolution kernel size of k; is the average value of all channel variances; G is the number of channels in each group; γ is the adjustment factor.

[0022] The following are the relevant experiments of the present invention

[0023] experiment:

[0024]

[0025]

[0026] Visual comparison:

[0027] To verify the effectiveness of ACAM, the study compared the initial YOLOv10n network and the YOLOv10+ACAM network using heatmap visualization. Figure 3-Figure 4 The results show that the ACAM attention mechanism can effectively solve the missed detection problem of the initial YOLOv10n network when facing defects such as insulator flashover discharge.

[0028] Comparison of detection effects:

[0029] To further verify the effectiveness of ACAM, the study used the YOLOv10n+ACAM model and the initial YOLOv10n model for a comparative test. Two pictures were randomly selected from the insulator discharge and insulator damage datasets as the input of the model detect. Figure 5-Figure 6 .

[0030] The above description is only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. A lightweight adaptive collaborative attention mechanism, characterized by: Specifically, the input features are divided into n different groups, each group is independently processed through global average pooling, adaptive convolution and sigmoid activation function to obtain the attention weight of the group, and then the attention weight is multiplied by the group of features to obtain the weighted features of the group, and finally the weighted groups of features are concat spliced ​​to generate the final output.

2. A lightweight adaptive collaborative attention mechanism according to claim 1, characterized in that: The size of the adaptive convolution kernel refers to the setting of the one-dimensional convolution kernel size in the ECA attention mechanism, and the average value of the variance of each set of feature channels is introduced on this basis. Let the convolution kernel size be dynamically adjusted according to the variance change of the input feature map.

3. A lightweight adaptive collaborative attention mechanism, characterized by: It includes the following calculation formula: Among them, X i is the input feature map of the i-th group, n is the number of feature groups; f k×k is a convolution operation with a convolution kernel size of k; is the average value of all channel variances; G is the number of channels in each group; γ is the adjustment factor.

Images