Research On YOLOv3 Model Compression Method And Its Application In Remote Sensing Image Object Detection

In recent years,remote sensing image object detection technology is increasingly becoming a hot research field of remote sensing,and widely used in areas such as satellite detection,disaster prevention management,urban planning,etc..However,these application fields have very high requirements for the accuracy and speed of object detection.For example,satellite monitoring needs not only to be able to detect the target in the image with more accuracy,but also to provide real-time feedback to the ground.Therefore,the speed and accuracy of object detection in remote sensing images have become the focus of attention.At present,the object detection technology based on deep learning has achieved good results in terms of accuracy,but the deep learning model has huge parameters,and requires huge computing resources,high memory space,and high-performance GPU to give full play to its advantages.Therefore,remote sensing object detection models based on deep learning cannot be deployed on fixed-performance and embedded platforms.For this problem,researchers have proposed many lightweight models and model compression methods.On the premise of ensuring the accuracy of object detection,they try to reduce the model parameters,the dependence of the model on hardware resources and calculation,and to improve the running speed of the model.This thesis mainly studies the compression methods of yolov3 model in the field of object detection,and establishes a lightweight remote sensing image object detection model.Under the premise of ensuring the accuracy as much as possible,we try to reduce the size and calculation of the model,and to improve the reasoning speed of the model.The main tasks completed in this thesis are as follows:(1)This thesis proposes an improved YOLOv3 channel pruning algorithm.Based on the network structure characteristics of YOLOv3,the limit pruning strategy in YOLOv3 channel pruning has the highest compression rate,but the number of channels of the pruned model may be not regular,which will bring bad influence to the effective use of memory resources and thus lower the inference speed of the model.This thesis proposes a new pruning algorithm to regularize the number of pruning channels in the limit pruning to 2~n.Although a small part of the compression rate is sacrificed,the model accuracy and inference speed are improved.(2)This thesis proposes a comprehensive YOLOv3 model compression method by fusing pruning,quantification and knowledge distillation together.The original YOLOv3 model is first pruned and compressed,then the network structure is quantified,and finally through knowledge distillation,the original YOLOv3 model transfers knowledge to the compressed model.This compressing method can reduce the size and complexity of the YOLOv3 model,and improve the reasoning speed of the model with the accuracy of the compressed model guaranteed.In this thesis,based on the DIOR remote sensing data set,we conduct objective analysis and comparison experiments in terms of model size,detection accuracy,inference time,and parameter amount of the new model compression method and the classic model compression method.Comprehensive experimental results show that the new model compression method proposed in this thesis is more effective.Finally,our compressed YOLOv3 model is applied to the online remote sensing object detection platform.In practical applications,the reliability of the compressed YOLOv3 model is verified.