Research On Remote Sensing Image Target Detection Algorithm Based On Deep Learning

With the rapid development of remote sensing and computer technology,a large number of high-resolution earth observation satellites have emerged,providing massive high-quality remote sensing data resources.Different from general images with relatively short shooting distances,remote sensing images have excellent characteristics such as a large field of view and strong data integration.They are widely used in environmental monitoring,resource exploration,regional planning and other fields.Remote sensing image is one of the main methods of earth observation.It is of great significance to study how to extract effective information from it quickly and accurately and use it.Due to its own characteristics,images often contain a large number of small targets and dense scenes.At the same time,there are problems such as target deformation and background indistinguishability.Therefore,accurate recognition of targets in remote sensing images requires the design of more targeted algorithms.In view of the fact that deep learning algorithms have performed well in natural image target detection tasks in recent years,they concentrate target feature extraction,classification and regression processes into neural networks,and increase the dimension of feature extraction,compared to traditional algorithms that rely on artificially designed features.The detection accuracy has been significantly improved.Among them,the method based on convolutional neural network shows high performance,which establishes the research direction of this article to improve the accuracy of remote sensing image target detection.This paper analyzes the current difficulties in remote sensing image target detection.Based on the two-stage target detection model Faster R-CNN framework,an improved model method is proposed according to actual needs.The main research work is as follows:(1)Proposed an anchor point supplement and dense scene optimization model based on FPN-Res Net.Aiming at the characteristics of remote sensing images,the residual network Res Net101 is used to replace the VGG-16 network to upgrade the core network of feature extraction.Further,for the problem that the small target features are not obvious after multiple feature extractions in the high-level network,which leads to the missed detection and classification errors of the small targets,the feature representation is enhanced by introducing the feature pyramid network to merge the advantages of different levels of feature maps.At the same time,the small target anchor point supplement is designed so that the small target anchor point can get more training opportunities and is not easily ignored.And the optimized non-maximum suppression method is adopted to allow the network to achieve a balance between eliminating duplicate frames and effectively avoiding missed frames.The targeted improvement of the model has significantly improved the detection accuracy of small targets in remote sensing images,and the detection effect in dense target scenes has also been improved.(2)A target detection model for multi-branch RPN remote sensing images based on deformable convolution is proposed.Research on target deformation and background interference problems that have not been solved before,improve the model’s ability to detect targets of various sizes.Introducing deformable convolution,using its deformability to make the convolution closer to the target shape,designing a multi-branch RPN structure,using different sizes of convolution kernels for the differences in the sensitivity of targets of different scales,and enhancing the network’s ability to recognize targets of various scales.The clustering algorithm is used to optimize the size of the target candidate frame,and the attention mechanism is integrated into the network to further select important information,suppress useless background information,and enhance the robustness of the network.Through experiments,the improved algorithm in this paper has a certain improvement in the detection effect of various types of targets on the public remote sensing data set,especially in complex background scenes.