Research On Small Target Detection Method Of Remote Sensing Image Based On Deep Learning

With the continuous progress of satellite remote sensing technology,researchers can obtain highprecision remote sensing images,which is of great significance to detect key targets.Due to the long shooting distance,remote sensing images contain a large number of small targets,such as aircraft,ships,and small cars.Such small targets occupy few pixels in the whole image,and not much information is available.Thus,their features are not easy to be extracted.This research aims at the problem of small target detection in remote sensing images,and proposes the improved D-YOLOv4 and YOLOv4-tiny AR small target detection algorithms based on YOLOv4.The contents of this thesis can be summarized as follows:(1)Firstly,the research background and significance of small target detection in remote sensing images are introduced,and the research status of traditional algorithms and small target detection is summarized.Then,this thesis introduces the principle and composition of each module of the convolutional neural network,and two common deep learning target detection algorithms are summarized.Besides,combined with the development trend,this thesis presents several lightweight backbone models and analyzes the limitations of general methods for small target detection.(2)Aiming at the small target detection in remote sensing images,an improved algorithm based on YOLOv4 is proposed.First,the Dense Net dense residual network is used to replace the backbone module of the original network.The dense connection method realizes the information interaction between different feature layers,and enables stronger information representation ability for small targets.Also,this research utilizes the depth-wise separable convolution,which effectively reduces the amount of parameters and computational complexity.The average detection accuracy of the improved D-YOLOv4 algorithm designed in this research reaches 92.78%,which is higher than the accuracy of the original by 4.87%,and the FPS reaches 26.(3)In order to further improve the detection speed,an improved algorithm based on YOLOv4-tiny is proposed.First,the CBAM attention module is added to the backbone part,and a new detection head for small targets is constructed after the first CSP module.Next,the concept of dilated convolution is introduced,and an improved RFBs module is constructed.Then,three effective feature layers are added to the constructed RFB-s module respectively,and this forms a new enhanced feature extraction network R-FPN together with the feature pyramid structure.Compared with the YOLOv4 algorithm,the parameter quantity of the improved YOLOv4-tiny AR target detection algorithm designed in this research is reduced by 56 M,the average detection accuracy reaches 92.71%,and the FPS reaches 65.Compared with the general network,the lightweight model detection efficiency is higher.