Research On Target Detection In Optical Remote Sensing Images Based On Deep Learning Algorith

Object detection of optical remote sensing image is a computer technology combined with computer vision and remote sensing technology.At present,optical remote sensing image object detection based on deep learning algorithm is always transformed from the traditional object detection model mostly,without considering the characteristics of remote sensing image,such as large range distribution of size,wide content,large amount of information and numerous small and dense targets.In addition,remote sensing image object detection often involves inclined object detection,which is more challenging than the horizontal object detection based on traditional model.In order to improve the detection accuracy of optical remote sensing image,based on the traditional detection model,this paper puts forward relevant solutions according to the characteristics of remote sensing image.The main work contents are as follows:(1)In terms of horizontal target detection,this paper proposes an improved algorithm based on Cascade R-CNN with multi-perception domain,which introduces HRnet and combines it with multi-perception domain and lightweight attention mechanism.In this paper,based on the cascade structure of Cascade R-CNN,a series-parallel Cascade R-CNN was designed to learn depth features multiple times at different scales.In the complex scene of multi-scale and multi-aspect ratio,the traditional algorithm has not enough ability to mine the deep features of multi-category objects,and the training strategies of many relevant algorithms are complex.In view of above problems,Our model can strengthen the mining and learning of the feature of complex remote sensing objects,without pretraining model and with simple training strategy.The average accuracy of the proposed model reaches 66.93% and 67.82% on the current two largest remote sensing image target detection datasets respectively.(2)In terms of inclined target detection,this paper proposes a remote sensing image target detection network based on feature fusion and weighted position regression loss.In this network,a feature pyramid network combined with multiple feature extractors is designed,which can enhance feature extraction and fusion and combine it into a target detection model named Oriented R-CNN.The improved model not only has the strong position regression ability of Oriented R-CNN,but also strengthens feature learning in different directions and multi-level feature fusion,to solve the problems of remote sensing image feature always has various directions and scales.In addition,this paper designs a weighted position regression loss function based on samples,which can dynamically adjust the weight of the loss function according to the position regression difficulty of different samples,so as to improve the accuracy of inclined target detection.Under the condition of single scale training and testing,this model achieves 77.3% average accuracy on DOTA dataset.