Research On Deep Convolutional Neural Network Based Object Detection Methods In High-resolution Remote Sensing Images

As an important task in the field of remote sensing image interpretation,target detection has always been one of the research hotspots in this field.With the development of imaging sensor technology,the resolution of remote sensing images is gradually improving,and the requirement for more and more refined target interpretation poses challenges to the accuracy and timeliness of detection and recognition of objects of interest in remote sensing images.Traditional target detection methods are mainly based on sliding window search strategies with hand-designed features or shallow features.These features have limited ability to express target characteristics,and methods based on sliding window search are inefficient.In recent years,target detection methods based on deep convolutional neural networks(Deep Convolutional Neural Network,DCNN)have made breakthrough progress in the field of computer vision,surpassing traditional methods in inference speed and detection accuracy,due to the rich details in optical remote sensing images and there is a certain similarity with the natural scene images.Therefore,the target detection methods based on DCNN is also widely used in the target detection of optical remote sensing images.However,there are differences in imaging methods,scales,backgrounds,etc.between remote sensing images and natural images,so when directly applying DCNN to remote sensing image target detection,there are several main challenges that still need to be faced:(1)In large-format remote sensing images,the background is more complex,the target size changes diversely,and at the same time,small targets such as aircraft and vehicles contain less information,it is difficult to accurately identify and locate them;(2)Because of the remote sensing images are top-view imaging,the direction of the target is arbitrarily distributed,and the direction of the target is very valuable for extraction;(3)In some scenes,the targets are densely distributed,and the boundaries between each other are blurred.For the several challenges in target detection of remote sensing images,this paper has carried out the research of DCNN-based optical remote sensing image target detection methods.Through the experiments and analysis on the public remote sensing image target detection data set,the effectiveness of the algorithm in this paper is verified.The main work done in this article is as follows:1.Designed and prepared the target detection data set of the optical remote sensing sewage treatment plant,analyzed and introduced the Faster R-CNN target detection framework based on the horizontal candidate area,and verified the shortcomings of the framework through experiments.For small targets,the framework is prone to missed inspections.2.In order to make the model adapt to the scale change of the target in the remote sensing image and improve the detection ability of small-size targets,an improved Faster R-CNN small-scale target detection method is proposed,which combines multi-layer and multi-scale feature maps.The features of small-sized targets are extracted on shallow feature maps with higher resolution,and the features of large-sized targets are extracted on deep feature maps with lower resolution.Finally,comparative experiments were carried out on a large-scale optical remote sensing image data set,and the generalization of the model was tested on a small-scale remote sensing image data set and a large-format optical remote sensing aircraft image data set to verify the effectiveness of the proposed method.3.Aiming at the problem of input sample distribution mismatch in the training and testing phases of target detection tasks,a target detection method based on cascade network is proposed.This network changes the distribution of input samples by increasing the threshold of positive and negative samples in the resampling process.To train a higher performance detector by using higher quality samples.In addition,for the random distribution of target directions in remote sensing images and the large difference in aspect ratio,a slanted frame cascade network was proposed and multiple types of target detection and direction extraction were performed on the public visible light remote sensing image target detection data set Experiments verify the effectiveness of the proposed target detection method.