Research On Object Detection Algorithm In Optical Remote Sensing Based On Deep Learning

Object detection is an important research direction in computer vision,which requires automatic detection of the position of the object of interest in the image or video and judgment of its category.In recent years,the rapid development of deep learning and high-resolution remote sensing technology has made it possible to accurately detect and identify geospatial targets.In terms of deep learning,Convolutional Neural Network(CNN)has become an important tool in the field of remote sensing image processing.Compared with traditional target detection algorithms-algorithms based on visual saliency,edge detection algorithms and shallow machines Learning algorithm,deep learning algorithm has great advantages;In terms of high-resolution optical remote sensing technology,with the rapid development of aerospace technology and high-resolution sensors,a large number of high-resolution optical remote sensing images with rich detailed information have been provided.Thanks to the rapid development of deep learning and remote sensing technology,the target detection technology in optical remote sensing images can provide reliable services for the civilian and military fields.In the civilian field,accurate and real-time remote sensing target detection can assist urban traffic management,improve fire detection efficiency,and help field rescues;in the military field,high-precision real-time remote sensing target detection can improve military reconnaissance efficiency and help accurately lock and track the military the goal.At present,target detection algorithms based on Deep learning can be divided into two categories: one is the two-stage detector represented by the R-CNN family,and the other is the one-stage detector represented by SSD and YOLO family.Both types of detectors have their own advantages and disadvantages,and their performance in different tasks is also very different.However,in general,the two-stage detector has a high detection accuracy and performs well in small target detection tasks,but the detection speed is not satisfactory.Compared with the two-stage target detector,the one-stage target detector has obvious advantages in detection speed,but the detection accuracy is not high,especially in a small target detection task with a high rate of missed detection.At the same time,these two kinds of detectors are proposed for the detection of conventional image targets at the beginning,and the effect of direct application to remote sensing targets detection is not good.Therefore,it is very necessary to explore the optical remote sensing target detector with both detection accuracy and detection speed.This paper analyzes the characteristics of remote sensing targets and constructs a remote sensing target data set ACS that includes three types of targets: airplanes,cars,and ships;for the complex remote sensing images,small remote sensing targets,large changes in target scale,etc.,the network features extraction of remote sensing targets.For the problem of poor ability,a Multi-scale Residual Block(MRB)is proposed,which uses jump connections and convolution with holes in the cascaded convolution structure to capture multi-scale context information and avoid the disappearance of gradients,thereby improving the model The feature extraction capability of remote sensing targets;the Multiscale Receptive Field Enhancement Module(MRFEM)is proposed,which combines the features obtained from MRB modules with different receptive fields to further enhance the multi-scale of remote sensing targets Feature representation capability;finally,based on the one-stage detector SSD,a Multi-vision Network(MVNet)is proposed to initially explore how to effectively use shallow convolution to detect small targets in remote sensing images in real-time,and use it in ACS,NWPU VHR-10 and RSOD three data sets and the performance of mainstream detectors have been comprehensively compared and analyzed.In summary,we constructed a remote sensing target data set containing three types of targets: airplanes,cars,and ships,and proposed a multi-scale residual block and a multi-scale receptive field enhancement module,and on this basis,we proposed a combination of detection accuracy and detection.Multi-vision detector for speed.The proposed multi-vision detector has achieved good detection accuracy and detection speed in the detection task of optical remote sensing targets and has a good guiding significance for target detection in optical remote sensing images.