Object Detection In Remote Sensing Images Based On Deep Learning

With the continuous development of modern remote sensing technology,a large number of images with higher spatial resolution and richer content have emerged,providing important analysis condition and resources for the research on various fields in remote sensing images.Object detection in remote sensing images,as the primary task of remote sensing image processing,is of great practical application value in military and civilian fields,and has been widely concerned and studied by scholars at home and abroad.In recent years,deep learning models,especially deep convolutional neural networks,have been widely and successfully applied to natural scene images object detection because of their excellent ability of semantic feature extraction.Due to the differences of the imaging modality between remote sensing images and natural scene images,when applying the deep learning methods to the field of remote sensing images directly,it will suffer from loss of small-sized objects?poor ability of anti-jamming and loss or misjudgment of dense objects and so on.Therefore,this paper researches on the application of deep learning methods to the object detection in remote sensing images,main work and contributions are summarized as follows:(1)To solve the problem that the current deep feature has no enough feature response to small-sized objects and poor ability of anti-jamming in remote sensing images,we propose a deep feature extraction method based on dilated convolutions and contextual information,and also embed it in the Faster R-CNN object detection framework,which enhances the response of the small-sized objects by improving the resolution of the feature map effectively,and adds context to assist the classifier to discriminate.In addition,due to the lack of open dataset resources of object detection in remote sensing images,we also construct and label the BODRS-2 and the TODRS-3 datasets for large ground objects and small ground objects respectively.Experimental results show that the proposed method can reduce the missing alarm rate and false alarm rate of the object detection in remote sensing images effectively.Moreover,it has high robustness and can obtain excellent detection performance on both small data sets and large data sets.(2)Most of the existing deep learning-based object detection algorithms utilize horizontal bounding box to locate objects,which causes inaccurate location of the objects with dense distribution and arbitrary direction in remote sensing images,thus leading to the missing detection.To tackle this issue,an object positioning method based on arbitrary-angle bounding box is proposed and embedded in the Faster R-CNN object detection framework.In this method,arbitrary-angle bounding box is introduced to locate objects with no redundancy.Also,In order to adapt to the ship objects with large length-width ratio,corresponding anchor ratios are added.In order to reduce the interference of the arbitrary-angle bounding box on the horizontal bounding box prediction,weight of the horizontal bounding box regression is increased.Compared to other classical object detection algorithms,the proposed method can reduce the missing alarm rate and false alarm rate of object detection in remote sensing images effectively.(3)In order to improve the efficiency of object detection,we propose a hierarchical object detection method for high-resolution satellite optical remote sensing images.Firstly,large ground objects are detected in the down-sampling image with low resolution,then regions of the detected objects is mapped to the original high-resolution image.Finally,the detection of small ground objects is carried out on the mapping regions.Based on the above work,we construct a test verification system for the remote sensing image object detection,and test and evaluate the performance of the hierarchical object detection method on the existing high-resolution satellite optical remote sensing image data.