CN-121999367-A - Lightweight remote sensing image target detection method and system suitable for edge equipment

CN121999367ACN 121999367 ACN121999367 ACN 121999367ACN-121999367-A

Abstract

The invention relates to the technical field of computer vision, and provides a lightweight remote sensing image target detection method and system suitable for edge equipment. The method comprises the steps of 1, obtaining a remote sensing image dataset and preprocessing the dataset to obtain a training set and a testing set, 2, constructing a remote sensing target detection model, wherein the remote sensing target detection model takes a YOLOv network as a reference model, introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module, 3, conducting end-to-end training on the remote sensing target detection model by utilizing the training set, and 4, inputting the testing set into the trained remote sensing target detection model, and outputting a target boundary frame and a category prediction result. The method and the system provided by the invention can improve the detection precision of the multi-scale targets in the remote sensing image and the actual deployment performance of the model.

Inventors

YOU DATAO
Lei Deyu
ZHAO BINGBING

Assignees

河南大学

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. The lightweight remote sensing image target detection method suitable for the edge equipment is characterized by comprising the following steps of: Step 1, acquiring a remote sensing image data set and preprocessing to obtain a training set and a testing set; Step 2, constructing a remote sensing target detection model, wherein the remote sensing target detection model takes a YOLOv network as a reference model, and introduces a feature selection aggregation downsampling module, an isomerism receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module, wherein the aggregation downsampling module is used for improving the expression capacity of shallow features, the isomerism receptive field feature enhancement module is used for enhancing the robustness and the direction consistency of feature expression, and the multi-branch multi-scale fusion attention mechanism module is used for improving the sensitivity and the robustness of the remote sensing target detection model to targets in a complex environment; Step 3, performing end-to-end training on the remote sensing target detection model by utilizing the training set; And 4, inputting the test set into a trained remote sensing target detection model, and outputting a target boundary box and a category prediction result.
2. The method for detecting the light-weight remote sensing image target suitable for the edge equipment according to claim 1, wherein in the step 1, the preprocessing comprises the steps of carrying out unified size normalization processing on the remote sensing images in the remote sensing image data set, carrying out data enhancement by adopting a plurality of online data enhancement strategies, and dividing a training set and a testing set according to a preset proportion, wherein the plurality of online data enhancement strategies comprise a Mosaic jigsaw, a color disturbance, random clipping and mirroring.
3. The method for detecting the lightweight remote sensing image target suitable for the edge equipment according to claim 1 is characterized in that in the step 2, a YOLOv network is used as a reference model in the remote sensing target detection model, a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module are introduced, and the method specifically comprises the steps that a CBS module and a C2F module are replaced by the feature selection aggregation downsampling module and the heterogeneous receptive field feature enhancement module in a backbone network of the YOLOv network respectively, and the multi-branch multi-scale fusion attention mechanism module is added at the joint of the neck network and the backbone network.
4. The method for detecting a lightweight remote sensing image target applicable to an edge device according to claim 1, wherein in step 2, the feature selection aggregate downsampling module is used for improving the expression capability of shallow features, and the specific processing flow is as follows: The input features of the feature selection aggregation downsampling module are subjected to 3×3 convolution processing to obtain first features; inputting the first characteristic into SPDConv module for processing to obtain a second characteristic; The first feature is processed by an expansion convolution block and SPDConv modules to obtain a third feature; The second feature and the third feature are spliced by a splicing unit and then respectively input into an average pooling block and a maximum pooling block to obtain a fourth feature and a fifth feature; processing the fourth feature and the fifth feature through a convolution block and a sigmoid function to obtain attention weight; And respectively carrying out channel point multiplication weighting on the second feature and the third feature by using the attention weight, and carrying out fusion and 1 multiplied by 1 convolution processing on the second feature and the third feature after the channel point multiplication weighting to obtain the output feature of the feature selection aggregation downsampling module.
5. The method for detecting the light-weight remote sensing image target suitable for the edge equipment according to claim 1, wherein in the step 2, the heterogeneous receptive field feature enhancement module is used for enhancing the robustness and the direction consistency of feature expression, and specifically comprises a3×3 convolution block, a parallel direction convolution branch and a dynamic weight adjustment branch, wherein the direction convolution branch comprises a first branch and a second branch, the first branch comprises a horizontal convolution block, a vertical convolution block and a1×1 convolution block which are sequentially connected, and the second branch comprises a vertical convolution block, a horizontal convolution block and a1×1 convolution block which are sequentially connected, and the dynamic weight adjustment branch comprises a3×3 convolution block, an activation function layer, a1×1 convolution block, a classification function and a weight value layer which are sequentially connected; And carrying out weighted fusion on the outputs of the direction convolution branch and the dynamic weight adjustment branch, and adding the weighted fusion with the input features of the heterogeneous receptive field feature enhancement module in a residual mode to obtain the output of the heterogeneous receptive field feature enhancement module.
6. The method for detecting the target of the lightweight remote sensing image suitable for the edge equipment according to claim 1, wherein in the step 2, the multi-branch multi-scale fusion attention mechanism module is used for improving the sensitivity and the robustness of the target detection model to the target in the complex environment, and concretely comprises a space attention branch and a multi-scale channel enhancement branch, wherein the space attention branch comprises an average pooling layer and a maximum pooling layer, a splicing layer, a1×1 convolution block, a3×3 convolution block and a sigmoid activation function which are sequentially connected in parallel, and the multi-scale channel enhancement branch comprises the average pooling layer, the 1×1 convolution block, the 3×3 convolution block and the 1×1 convolution block which are sequentially connected; and after the outputs of the spatial attention branches and the multi-scale channel enhancement branches are weighted element by element, adding the spatial attention branches and the input features of the multi-branch multi-scale fusion attention mechanism module through residual connection to obtain the output of the multi-branch multi-scale fusion attention mechanism module.
7. The method for detecting the lightweight remote sensing image target suitable for the edge equipment according to claim 1 is characterized by further comprising the steps of carrying out preliminary screening on detection frames output by the remote sensing target detection model according to a preset confidence threshold value to obtain candidate detection frames, carrying out non-maximum suppression operation on the candidate detection frames, calculating the overlapping degree among the candidate detection frames, suppressing redundant overlapping frames in the same target area, and reserving a boundary frame with the highest confidence as a target boundary frame.
8. A lightweight remote sensing target detection system for edge devices, comprising: the data set acquisition unit is used for acquiring a remote sensing image data set and preprocessing the remote sensing image data set to obtain a training set and a testing set; The remote sensing target detection model takes a YOLOv network as a reference model, and introduces a feature selection aggregation downsampling module, an isomerism receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module, wherein the aggregation downsampling module is used for improving the expression capacity of shallow features, the isomerism receptive field feature enhancement module is used for enhancing the robustness and the direction consistency of feature expression, and the multi-branch multi-scale fusion attention mechanism module is used for improving the sensitivity and the robustness of the remote sensing target detection model to targets in a complex environment; the model training unit is used for carrying out end-to-end training on the remote sensing target detection model by utilizing the training set; and the target detection unit is used for inputting the test set into the trained remote sensing target detection model and outputting a target boundary box and a category prediction result.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1 to 7 when the program is executed by the processor.
10. A non-transitory computer readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the method of any one of claims 1 to 7.

Description

Lightweight remote sensing image target detection method and system suitable for edge equipment Technical Field The invention relates to the technical field of computer vision, in particular to a lightweight remote sensing image target detection method and system suitable for edge equipment. Background Along with the rapid development of remote sensing imaging technology and high-resolution satellite images, remote sensing image target detection is increasingly widely applied in the fields of homeland resource investigation, city management, environment monitoring and the like. The remote sensing image often contains abundant ground object information, such as buildings, vehicles, planes, ships and the like, and the target detection is one of key technologies in the intelligent analysis of the remote sensing image, so that the method has important significance for realizing automatic information extraction. However, compared with natural images, the Optical Remote sensing image (Optical Remote SENSING IMAGES, ORSI) has the characteristics of more complex background, changeable scale difference, weak semantic consistency and the like, and the characteristics significantly improve the technical challenges of the target detection task. Current remote sensing image target detection techniques still face a number of challenges. Firstly, the background of the remote sensing image is complex and changeable, and interference elements such as buildings, vegetation, shadows and the like are often mixed with foreground targets, so that the detection model is difficult to accurately separate the targets from the background under the condition of lacking explicit guidance or auxiliary supervision, and the detection precision is reduced. Secondly, the obvious difference of the multi-scale targets also puts higher requirements on the detection capability, and particularly in a small target detection task, the problems of small target size, weak semantic expression, frequently-occurring characteristic suppression, boundary blurring and the like are easy to cause missed detection and false detection. In addition, in order to improve the detection performance, a large number of complex modules are introduced into a part of methods, such as a multi-layer pyramid structure, a concentration mechanism combination, auxiliary branches and the like, and the redundant structures obviously increase the calculation burden and the parameter scale of a model although improving the precision, so that the real-time deployment on resource-limited equipment such as an unmanned plane, a satellite terminal and the like is not facilitated. Finally, the existing feature extraction mode still has limitation on expression capability, although high-level semantic features have good semantic abstraction capability, spatial detail information is sacrificed, while low-level features have fine edge expression, but semantics are incomplete, and the existing fusion strategy is difficult to fully combine the advantages of the high-level features and the low-level features, so that the further improvement of the overall detection performance is restricted. Disclosure of Invention Aiming at the problems of complex background interference, insufficient small target detection capability and insufficient target detection accuracy caused by model structure redundancy in the existing remote sensing image target detection method, the invention provides a lightweight remote sensing image target detection method and system suitable for edge equipment, and the detection accuracy of a multi-scale target in a remote sensing image and the actual deployment performance of a model can be improved. In a first aspect, the present invention provides a lightweight remote sensing image target detection method applicable to an edge device, including: Step 1, acquiring a remote sensing image data set and preprocessing to obtain a training set and a testing set; Step 2, constructing a remote sensing target detection model, wherein the remote sensing target detection model takes a YOLOv network as a reference model, and introduces a feature selection aggregation downsampling module, an isomerism receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module, wherein the aggregation downsampling module is used for improving the expression capacity of shallow features, the isomerism receptive field feature enhancement module is used for enhancing the robustness and the direction consistency of feature expression, and the multi-branch multi-scale fusion attention mechanism module is used for improving the sensitivity and the robustness of the remote sensing target detection model to targets in a complex environment; Step 3, performing end-to-end training on the remote sensing target detection model by utilizing the training set; And 4, inputting the test set into a trained remote sensing target detection model, and outputting a target bo