Learning-based Object Detection Methods For Remote Sensing Images

Different from normal images which are captured at short range,remote sensing images are of large scale and macroscopic view,so object detection for high-resolution remote sensing images has been widely used in numerous scenes such as life and industry.Since there are high shooting height,wide coverage,and a large number of small objects and dense scenes in high-resolution remote sensing images,accurate object detection for remote sensing images needs some more targeted algorithms.Deep learning methods for object detection unify the processes of target feature extraction,classification and position regression into a single neural network structure,which makes the feature dimension much higher than that obtained by manual feature extraction methods.Among learning-based object detection methods,the convolutional neural network(CNN)approach can usually achieve high performances in object detection and has been widely used in image processing.This thesis proposes a CNN-based model for object detection in high-resolution remote sensing images.Based on the standard Faster R-CNN detection framework,the proposed model uses the methods of feature enhancement and target saliency to solve the problem of low precision of small object detection in remote sensing images.Further,the attenuated nonmaximum suppression method is used to solve missing detection problems of dense scenes in remote sensing images.The proposed model is mainly divided into three parts.Firstly,the feature enhancement structure adopts sharpening convolution kernels combined with convolution regularization to enhance the underlying detailed features,thereby improving the detection accuracy of small objects.Secondly,the target saliency structure abandons the standard RoI pooling process in the Faster R-CNN framework with target up-sampling and re-convolution methods to solve the feature losses and mosaic problems caused by RoI pooling,especially for the detection accuracy of small objects.In addition,the attenuated non-maximum suppression structure is used to optimize the standard non-maximum suppression process,and balances between avoiding effective targets missing problems and eliminating duplicated detection targets,so that the detection accuracy in dense scenes can be effectively improved.Based on the open source large-scale high-resolution remote sensing dataset DOTA,the experimental results show that the proposed model performs better for small objects in remote sensing images and has better effects for detection in dense scenes.The proposed model has an average accuracy of 6.1%,19.32%,27.36%,and 55.62% improvement compared to the standard Faster R-CNN,R-FCN,YOLOv2,and SSD algorithms.