Research And Implementation Of Object Detection And Recognition Algorithm For Optical Remote Sensing Image

With the rapid development of aerospace technology,satellite optical remote sensing technology has made great progress.Not only has the image resolution been greatly improved,but also has the ability to detect and transmit data in real time throughout the day,making the optical remote sensing image more efficient.The amount of data and the amount of information contained in a single image has grown rapidly.The rich information in the massive optical remote sensing images contains huge value,and it has a wide range of applications in all fields of society.How to extract valuable information efficiently and accurately has become a research hotspot in the field of computer vision.The object information contained in the optical remote sensing image is a very important part of the value of the optical remote sensing image.This paper mainly focuses on the research work of object detection and recognition technology for optical remote sensing images.The main work of the paper is as follows:(1)Collect and organize relevant data sets.Based on the Dota data set,the original image was cut to generate more than 10,000 images of the same size.At the same time,a script was written to convert the original 8-dimensional parameter annotation data to produce optical remote sensing image horizontal bounding box and rotating bounding box object detection data sets.(2)Through experimental comparison,find a benchmark algorithm for the horizontal bounding box object detection of optical remote sensing images,and improve the algorithm.Based on the existing excellent models such as YOLOv4,SSD and Retina Net,the prepared data set was used to train a horizontal bounding box object detection model for optical remote sensing images,and the model was tested.Through experimental comparison,it is found that YOLOv4 has the best detection effect,with m AP reaching 75.15%,which is much higher than the latter two.Improve the YOLOv4 model based on the defects of the horizontal bounding box object detection in the optical remote sensing image,and propose a method to set the NMS threshold value by category,which improves the model m AP by 2.53%;research the rationality of the anchor assignment of the YOLOv4 model and try several improvements design,in which the method of subscale clustering of anchor improves the detection performance of some object categories.(3)Designed and completed the rotating bounding box object detection algorithm based on YOLOv4 model.In order to obtain a bounding box with a higher degree of fit to the object and solve the problem of overlapping bounding boxes in the detection of horizontal bounding boxes,a more accurate rotating bounding box is derived.Improve the representation of the long side of the rotating box.Based on the YOLOv4 model,three rotating bounding box object detection models are designed and implemented: RYOLOv4 direct regression angle,RYOLOv4 angle segmentation,and RYOLOv4 multi-angle anchor model.The test results show that the performance in terms of RYOLOv4 multi-angle anchor model detection effect is the best,and the m AP is 65.3%.(4)Developed an optical remote sensing object detection system.The system can process data such as pictures and videos,and is designed with an image cutting detection function,so that the system can process large-scale image data such as remote sensing images.