Research On Arbitrary-Oriented Object Detection Algorithm In Remote Images Based On Deep Learning

In recent years,breakthroughs in remote sensing technology have been made,and a large number of high-resolution remote sensing images have been acquired.Currently,the use of deep learning techniques in remote sensing image target detection tasks has made great progress and is widely used in different scenarios.The use of deep learning models represented by convolutional neural networks can effectively free up human resources and leave the feature extraction to the convolutional algorithm,which greatly improves the robustness of the network for different data.we propose an anchor-free baseline network for remote sensing image arbitrary-oriented object detection based on the existing detection algorithms,which is oriented to the problems faced by remote sensing image target detection in the field of deep learning.In view of the difficulties of remote sensing image target scale change and target direction uncertainty,the performance advantages of adaptive spatial feature fusion and convergent path feature pyramid network are combined,and the convergent path adaptive feature pyramid target detection network is proposed,and Mobile Net-V2,Res Ne Xt-50 and Res Ne Xt-101 are used as the backbone networks for testing respectively.To address the difficulties such as complex background of remote sensing images,large variation of target scales and insufficient samples,this paper proposes a backbone network NLARes Ne Xt based on an improved non-local attention mechanism,and constructs a new backbone network with an FPN with improved upsampling operation.The improved attention mechanism uses grouped convolution to reduce the computational effort of the network.In this paper,comprehensive experiments are conducted on the DOTA dataset and the HRSC2016 dataset.The experimental results show that the method designed in this paper has a high accuracy rate,verifying the effectiveness and feasibility of the network proposed in this paper.