Study And Implementation Of Remote Sensing Image Object Detection Based On Deep Learning

High-precision object detection has always been a research hotspot in the field of remote sensing.It is difficult to recognize complicated targets on remote sensing images simply by relying on fixed mathematical mechanisms.With the development of deep learning,there are better solutions for image object recognition.Due to the powerful learning ability and versatility of neural networks,the structure proposed in a certain field can be easily transferred to other fields.As a typical object detection algorithm,YOLOv3(You Only Look Once v3)algorithm has the advantages of fast detection speed,simple structure and high detection accuracy.Because there are many small targets in remote sensing images,the separation of the target and the background is low,and the target is deformed,the detection effect of the YOLOv3 algorithm is satisfactory.And the data enhancement method of YOLOv3 is relatively backward.The YOLOv3 algorithm has a higher demand for computing power.It is difficult to deploy on edge computing devices such as satellites and drones.Aiming at the problems of the YOLOv3 algorithm,this thesis improves the detection effect of the YOLOv3 algorithm on remote sensing images and reduces the number of YOLOv3 parameters.The improvement can be divided into two aspects: optimizing the training process and improving the structure of the algorithm.This thesis uses mosaic data enhancement to replace the multi-scale training in YOLOv3 to improve the generalization.This thesis uses CIo U(Complete-Io U)instead of Io U(Intersection Over Union)to calculate the loss of bounding box regression to improve the detection accuracy of YOLOv3.In terms of improving the network structure,this thesis adds the spatial pyramid pooling layer to the feature pyramid network to improve the detection accuracy of the network for small targets,and replace part of the convolutional layers in the backbone network with grouped convolution to reduce the amount of network model’s parameters.In this thesis,the comprehensive performance and calculation amount are considered to experiment on the number of groups,and the YOLOv3-g2 model is selected as the final model of this thesis.Finally,this thesis uses transfer learning to improve the generalization of the YOLOv3-g2 model.This thesis is experimentally verified on the DIOR(object Detection in Optical Remote sensing images)data set.After using mosaic data enhancement,this thesis successfully increased the detection effect of the YOLOv3 algorithm by 12%.After using CIo U to optimize the frame regression process,the detection effect of the YOLOv3 algorithm is increased by 0.4%.Compared with YOLOv3,the YOLOv3-g2 model successfully reduces the amount of parameters and calculations by more than 50%,and the detection effect remains unchanged.Finally,this thesis uses transfer learning to increase the m AP@0.5 index of the YOLOv3-g2 model on the DIOR data set to 71.7%,far exceeding YOLOv3 and other detection algorithms.