Research On Remote Sensing Image Architectural Target Detection Method Based On Deep Learning

With the rapid development of remote sensing technology,remote sensing technology has been widely used in many fields such as meteorological inspections,resource census,agricultural production,and military applications.As one of the most widely distributed targets in human activities in remote sensing images,building target detection is of great value for urban planning,safe production,and research on regional social development.Facing the traditional target detection operator based on a fixed feature in inadequate in complex scenarios,based on the target detection depth study has obvious advantages in remote sensing image processing.Based on the existing target detection algorithm convolutional neural network,one is based on two-stage process FasterR-CNN represented,such methods known to detect a high accuracy,but its detection speed is slow;the other is YOLO,SSD is a representative one-stage method.This type of method is known for its fast detection speed,and its accuracy is slightly lower than that of the two-stage method at the same time.In recent years,the YOLO series of algorithms have developed rapidly,and their accuracy and speed have been greatly improved.In this paper,launched in April 2020 YOLO v4 is based on research and implement remote sensing image building target detection task:(1)The three important processes of the backbone network,feature extraction and fusion,target determination and positioning in the YOLO v4 algorithm are sorted out.And suggest improvements in targeting both the stage and the backbone network algorithms: in predicting the frame positioning,presented an analysis based on K-means ++ algorithm clustering DIUO value,revised forecast is unable to determine the target frame and the frame of the real time calculation of IOU and the distance between the degree of coincidence problems;in the backbone network,the original network CSPDark Net53 modify lightweight Mobile Net model,and the model suitable for the crop.Due to the addition of the deep separable convolutional network and the inverse residual,the multi-adds amount and model size are greatly decreased while maintaining the level of accuracy,and the model can be compressed to 1/9 of the original under ideal conditions.(2)Construct a data set,experiment and analyze.This article uses the WHU building dataset as training and test data,and the original data is re-text-labeled in the experiment.Experiments were carried out in two subsets,and the improved experiments were improved to different degrees in the two data set subsets: for the positioning phase improvement,the m AP of the improved experiment was increased by up to 1.6%,and for the backbone network,the m AP was increased by up to 2.5%,Combining the above two improvements,m AP can be increased by up to 3.3%,while reducing the model volume to 1/5 of the original size.Comparing the index parameters of each experiment and the detection effect diagrams under several typical backgrounds,the improvement of the improved experimental performance can be observed.