Research On Remote Sensing Image Target Detection Algorithm Based On YOLO

In recent years,with the continuous improvement of remote sensing technology,a large number of imaging satellites have emerged.Remote sensing images have attracted attention in many fields.Remote sensing images target detection has been widely used in many fields such as national defense security,urban construction planning,and disaster monitoring.Because the traditional target detection algorithm can not avoid the artificial design features in the feature extraction stage,it affects the performance of the target detection to some extent,and it is difficult to effectively analyze and utilize the remote sensing data.In the context of big data and automatic identification,with the deepening of research on deep learning algorithms in recent years,how to use remote sensing data and deep learning methods to detect remote sensing image targets from data to knowledge and improve remote sensing The efficiency of data utilization and the effectiveness of remote sensing applications are of great significance.Therefore,this paper is based on the YOLO algorithm with obvious advantages in the deep learning target detection algorithm.The algorithm is optimized for missed detection and false detection in the detection of different sizes,length and width ratios and small size objects.Based on the version of YOLOv3,an extended YOLO model called "Dliated-YOLO" is constructed and applied to remote sensing image target detection.The specific work is as follows:(1)For the single-stage model represented by YOLO,the shortcomings of target missed detection and small target detection rate.Three typical small target data of “bridge,seaport and airport” in remote sensing image are collected,At the same time the images was augmented by rotating,scaling,increasing salt and pepper noise and Gaussian noise as the sample training set and validation data set for this experiment.(2)Secondly,using the K-Means algorithm to cluster the remote sensing datasets of this paper,the anchor point dimension suitable for remote sensing data is obtained.(3)Furthermore,in view of the disadvantage of YOLO for small target detection,the dilated convolution is introduced into the feature extraction layer of Darknet53,and the local receptive field of the convolution kernel is expanded without increasing the parameter amount,Improving detection rate of the small target.At the same time,the model depth is further deepened,and the original three scale detections are increased to four scales.Finally,an extended YOLO model for the remote sensing image target detection,namely the Dilated-YOLO model,is constructed.(4)Finally,the Dilated-YOLO model constructed in this paper is applied to remote sensing image target detection.The experimental results show that the algorithm has a significant improvement in the small target recall rate,and the false detection rate is also controlled.At the same time,the convergence of the algorithm is also very good.The theoretical and experimental effects show that the model constructed in this paper can effectively improve target detection performance.