Research On Target Detection Algorithm Of Remote Sensing Image Based On YOLOv3

With the popularization and development of remote sensing image acquisition technology,high-resolution aerial imagery is becoming more and more common around the world.It contains a large number of information that can be associated with land development,maintenance,disease control,defect location,military surveillance and other applications.Target detection in remote sensing images has important reference significance for many tasks such as image understanding and target tracking in video,so it has high research value.At present,target detection tasks are roughly divided into two methods based on traditional methods and deep learning.Among them,deep learning-based methods are divided into two-stage and single-stage target detection methods.Considering the problem that artificial design features of traditional method have narrow range of applicable,weak generalization ability,and the slow detection speed of the two-stage target detection method of deep learning,this paper uses a single-stage target detection method based on deep learning.The main work is divided into three parts.First,this paper uses the YOLOv3 algorithm that can guarantee a good balance in detection speed and detection accuracy in the mainstream detection framework,and uses the DOTA remote sensing dataset to perform target detection tasks.In order to improve the applicability and accuracy of the algorithm,this paper improves the YOLOv3 algorithm.The K-Means method is used to perform dimensional clustering on remote sensing dataset labels,calculate the optimal width and height values,and modify the anchor value in the original YOLOv3 algorithm.Aiming at the different sizes of the instances in the remote sensing dataset,and the disappearance of small target receptive field after multiple convolutions,the network structure of the YOLOv3 algorithm is improved,and the SPP module and the ResConv module are added.Under the premise of ensuring the detection speed,the convergence speed is accelerated,and the detection accuracy of the algorithm is improved by 5.3%.Then,aiming at the problems of large image size,small target proportion,and many small targets in dense remote sensing datasets,this paper proposes the YOLO-SR-Attention algorithm,which introduces the attention mechanism based on the previous work.Adding an attention module to the feature extraction network Dark Net-53,which is the backbone network,enables the YOLO-SR-Attention algorithm to select information that is more critical to the current task target from a large amount of information.This improvement greatly improves the detection accuracy of small targets in aerial data sets.Among them,the detection accuracy of typical small target ships increased significantly,from 42.7% to 73.3%.In the overall data set,the detection accuracy reached 71.3%.Finally,in order to solve the problems of the improved YOLO-SR-Attention network with more parameters,large memory usage,and difficult to use on hardware devices with weak computing capabilities,this paper performs model compression and pruning to reduce unnecessary channel.Compressing it into a lightweight network while ensuring that the detection accuracy is almost the same.