Research On Aerial Remote Sensing Image Target Detection Based On Convolutional Neural Network

With the development of the aerospace industry and the advancement of observation technology,the role of aerial remote sensing images in the military and civilian fields is becoming more and more important.The characteristics of aerial remote sensing images are wide angle of view,many targets,and complicated background.At the same time,aircraft and ships are two of the more important targets in aerial remote sensing images.The target detection research work in this paper is based on these two targets.In this paper,based on the DOTA dataset,the datasets MyPlane and MyShip are used for network model training and detection.Aircraft and ships are small targets in aerial remote sensing images,which brings difficulty and challenge to target detection.The data processing results obtained only by manual interpretation or traditional target detection technology lack accuracy and timeliness.The detection of small targets under complex background still has the problems of low accuracy and slow speed.In order to solve these problems,this paper proposes an improved network model for improving and optimizing the YOLOv3 network model from three aspects:increasing the detection scale,rationally setting anchor frame parameters,and adding L2 regularization constraints in the loss function.First,add 104*104 detection scales to the three existing detection scales of YOLOv3,in order to enhance the sensitivity and detection ability of the network to small targets.Then,the K-means clustering algorithm is used to perform clustering operations on the two aerial remote sensing data sets respectively.Through clustering analysis,a reasonable number of anchor frames and anchor frame sizes are set in the improved model to enhance the network model.The ability to detect aircraft targets or ship targets.Finally,the overall loss function of the improved model in this paper adds L2 regularization constraints on the basis of YOLOv3 to avoid overfitting.Finally,this paper completes the training and testing of Faster R-CNN+Resnet101,Faster R-CNN+VGG16,YOLOv3 and the improved model of this paper on the MyPlane dataset and MyShip dataset,and analyzes the experimental results.The improved model based on YOLOv3 proposed in this paper performs well on two data sets.On the MyPlane dataset,the improved model has a 2.43%higher accuracy rate than YOLOv3,a 1.08%recall rate,and an average AP accuracy of 2.69%.On the MyShip dataset,the improved model has improved the accuracy rate by 0.37%compared to YOLOv3,the recall rate has increased by 1.61%,and the average accuracy AP has been increased by 1.81%.Compared with YOLOv3,Faster R-CNN+Resnet101 and Faster R-CNN+VGG16,regardless of the index parameters or the actual detection result graph,the network synthesis capability of the improved model in this paper has been improved.At the same time,the sensitivity of the improved model in this paper to the detection of small targets has been greatly enhanced.