Research On Remote Sensing Image Object Detection Based On Deep Learning

With the continuous development of modern remote sensing technology,the resolution of remote sensing images is becoming higher and higher,yielding massive amounts of information and more pressure to remote sensing image interpretation.Object detection is not only an important task in remote sensing image interpretation technology,but also a tough research topic in the field of computer vision.Recently,deep learning technique is developing rapidly and has made great progress in image classification and object detection in the field of computer vision,which therefore provides an efficient candidate to solve the shortcomings of inefficiency and high complexity caused by highly dependent on artificial design features of traditional remote sensing object detection algorithms.Therefore,it is of great significance to explore deep learning based object detection technique in remote sensing images.Compared with natural images object detection,remote sensing object detection faces more tough issues,for example,the objects in remote sensing images have the characteristics of multi-scale and arbitrary direction distribution,and there are many small objects with densely distribution,which brings challages to the object detection in remote sensing.Aiming at multi-scale,arbitrary direction distribution,small objects,dense distribution,complex background,and limited number of samples problems in remote sensing images,based on the deep learning,this paper focuses on multi-scale object detection,oriented bounding boxes(OBB),and efficient few-shot meta-learning based object detection.Specifically,the main research contribution and results are summarized as follows:(1)Aiming at the problem of multi-scale objects in remote sensing images that the convolutional neural network model is easy to lose the characteristics of small objects,a CSP-Hourglass Net structure with densely connected up and down sampling is proposed and extended in the YOLOv3 framework.Simulation results show that the proposed network can improve the performance of the algorithm in small object detection,and further enhance the visual detection effect by improving the non-maximum suppression method.(2)Aiming at the problems of multi-small objects and dense distribution resulting in unclear visual detection in remote sensing images,an OBB detection method based on vertex coordinate regression is proposed.It is embedded in the Center Net framework.By abandoning the general angle regression method,and detecting the OBB of the remote sensing image the presentation is enhanced.Furthermore,to deal with the problems of the arbitrariness of the object distribution and the complexity of the background,a random rotation mosaic data augmentation is proposed.Simulation results show that the proposed vertex regression method has better detection performance than the angle regression-based algorithm,and can represent the target more accurately and compactly.In addition,the proposed method can improve the detection accuracy of the algorithm.(3)In view of the fact that the number of images in the remote sensing image data set is far less than that of the natural image data set,and the poor generalization ability of the algorithm.A remote sensing object detection algorithm based on the meta-learning MAML framework is proposed.The proposed algorithm integrates the Center Net framework and the meta-learning framework,and implements effective few-shot object detection.Simulation results show that the proposed algorithm greatly reduces the dependence on the number of samples,and can achieve rapid convergence and effective detection for new tasks.