Research On Object Detection Algorithm Of Remote Sensing Image Based On Deep Learning

With the development of satellite technology,the resolution of remote sensing image gradually increases and the content contained in the image becomes more and more abundant,which promotes the development of remote sensing image processing.Remote sensing image target detection is of great significance and broad application prospects in both military and civilian fields.At present,there are many object detection methods,among which deep learning has achieved good results in the field of object detection and has great potential in remote sensing image detection.Therefore,deep learning object detection algorithm YOLOv4 was improved in this paper and applied to remote sensing image object detection.The specific work is as follows:First of all,when using YOLOv4 algorithm to directly detect objects in remote sensing images,there was a high rate of missing small objects.The YOLOv4 algorithm was optimized and improved from the two aspects of initial anchor frame size and feature fusion.According to the scale distribution characteristics of objects in the remote sensing image data set DOTA,the initial anchor frame was reset by using the K-means++clustering algorithm.Meanwhile,the method of adding high-resolution feature layer was used to improve the YOLOv4 network.The improved algorithm YOLOv4-F with improved detection accuracy was finally obtained.Secondly,aiming at the problem that remote sensing image object detection is prone to false detection and missing detection under complex background interference,YOLOv4-F algorithm is further improved and a object detection algorithm based on fusion semantic segmentation is proposed.While the YOLOv4-F algorithm was used for object detection,the VGG-Unet network was used for semantic segmentation of the image to obtain pixel-level classification information.According to the objective relation between the detected object and the background,the semantic classification information was used to remove the unreasonable detection results.The YOLOv4-UF algorithm with a reduced false detection rate was finally obtained.Finally,the traditional YOLOv4 algorithm was transplanted to the embedded development platform of NVIDIA Jetson TX1 for the training and testing of detection models.The remote sensing image object detection and processing of this algorithm was realized on the embedded development platform.In addition,the YOLOv3 algorithm was used for comparative experiments,proving that YOLOv4 algorithm was more adaptive to embedded remote sensing image target detection.