Research And Implementation Of Object Detection Technology Based On Deep Learning For Remote Sensing Image

Object detection has always been an important and basic research direction in the field of computer vision,and there is also a huge application demand in the industry.In recent years,general object detection algorithms have developed rapidly,and the detection accuracy and detection speed have been greatly improved.However,the object detection algorithm for remote sensing images is still difficult to achieve the detection effect of natural images in terms of detection accuracy and speed.Therefore,how to accurately describe the location information of the target and avoid the overlapping and interference of the targets is the key to improving the object detection effect of remote sensing images.Although the horizontal detection bounding box of the general object detection algorithm can accurately describe the position information of the target in the natural image,it is not effective for describing the target of the remote sensing image.According to the characteristics and challenges of remote sensing images,this thesis proposes corresponding improvement methods.The main work is as follows:（1）In view of the characteristics of random oriented angle and dense distribution of object in remote sensing images,this thesis proposes a object detection algorithm based on the oriented bounding box,and replaces the horizontal detection bounding box with the Midpoint Offset Representation to solve the problem of inaccurate target information representation..The experimental results show that the improved model based on the rotation detection frame proposed in this thesis significantly improves the prediction accuracy.（2）In view of the characteristics of random oriented angle and incomplete features of object in remote sensing images,this thesis proposes a data augmentation algorithm named Rotated Mosaic Augmentation.By rotating and splicing one origin image and other three random images,the feature of direction and local feature missing of the training data set can be greatly enriched.The experimental results show that the Rotated Mosaic Augmentation algorithm proposed in this thesis can improve the robustness and accuracy of the model.（3）By integrating the current rich remote sensing image data sources,effective object detection models,and out-of-the-box available Internet development technologies,this thesis constructs a remote sensing image object detection system that combines theoretical research and application practice.The algorithm in this thesis does experiments and compares accuracy mainly on the DOTA dataset,demonstrating that the improved YOLOv5 model has better robustness and generalization.