Object Detection In Remote Sensing Based With Deep Convolutional Neural Networks

The detection and recognition of ships in high-resolution remote sensing images play a vital role in the understanding of remote sensing images and is the basis of many remote sensing image analysis tasks.With the development of information and remote sensing technology,the optical remote-sensing images with higher spatial resolution make it possible to detect and identify remote sensing objects precisely.In recent years,deep learning method has made a breakthrough in the field of natural scene object detection.Both in detection accuracy and algorithm robustness,it surpasses the method based on traditional manual design features.However,due to the characteristics of objects in high-resolution remote sensing images(such as viewing angle of images,distribution of objects,various scale of targets,etc.),simply applying the detection algorithm for natural images to remote sensing images can hardly obtain satisfactory performance.This paper mainly aims at the problem of ship detection in high-resolution optical remote sensing images:(1)A remote sensing ship detection algorithm based on rotating rectangular region is proposed.Firstly,the rotated rectangular box is used to represent the detection results of ship targets.Through analysis,this representation method is superior to the current horizontal holding box representation method in terms of visual perception and algorithm performance.In order to realize ship detection based on rotating rectangular box,RPN network is improved based on the Faster RCNN algorithm to output candidate regions with rotating angles.(2)Proposed the pyramidal pooling module of regional features of interest.Large scale variation is a difficult problem in remote sensing target detection.In this paper,a pyramidal pooling module of interest region features that can integrate multi-scale pooling features is proposed to extract more effective image features of interest region,so as to complete better classification and improve the ability of algorithm to process multi-scale targets.(3)The positioning accuracy prediction branch is designed,and the post-processing algorithm is optimized by introducing the positioning accuracy to guide the non-maximum suppression algorithm.Currently,the classification confidence is used as the sorting basis of the target boxes in the post-processing process,but according to the analysis,it is found that there is a mismatch between the classification confidence and the quality of network positioning results.Therefore,the positioning accuracy prediction branch was designed,and the quality of the positioning results was scored by regression network.In the non-maximum suppression algorithm,the positioning score was used as the sorting basis of the target box,and the score of the prediction box was updated in the algorithm process to optimize the post-processing results.Experiments on HRSC2016 dataset show that proposed method outperforms exiting methods.