Research On Object Detection Algorithm Based On Deep Learning

With the advent of the era of big data,target detection algorithms based on deep learning have greatly improved detection accuracy and real-time performance,and have played a major role in security,smart construction sites,autonomous vehicles,and so on However,today's target detection algorithms have the problems of large model storage,high computing resources,and magnitude parameters.It is difficult to implement on platforms with low computing resources and small storage capacity.Lightweight algorithm models with high real-time and high accuracy are still to be studied.Therefore,a lightweight target detection algorithm is a research direction todayIn order to reduce the size of the model and increase the speed of the model,this paper proposes to use a lightweight network MobileNetv2 and YOLOv3 to obtain a new YOLOv3-MobileNetv2 network architecture,based on the original YOLOv3 algorithm,which is very fast After fusion,the detection speed of the network has been greatly improved,and the size of the trained model has been significantly reduced.At the same time,in order to further improve the detection accuracy of YOLOv3-MobileNetv2,this paper proposes a method based on the new evaluation index GIOU to improve its loss function,and the effect of the model has been improved again.The following is the main research content and work of this article:(1)This article first describes the target detection algorithm using artificial features and the target detection algorithm using deep learning,then explains the various components of the convolutional neural network,and then introduces the basic knowledge and structure of the target detection algorithm.YOLOv3,a mainstream fast target detection algorithm,conducts research and analyzes its algorithm principle and network architecture in detail(2)This paper proposes a brand-new YOLOv3-MobileNetv2 network architecture By integrating the lightweight network MobileNetv2 and YOLOv3,the model parameters and calculation costs are greatly reduced.At the same time,the network uses pyramid-like multi-scale features for fusion,which significantly improves the detection effect of targets of different sizes.The algorithm is trained on two public target detection datasets PASCAL VOC and MS COCO.The experimental results show that the improved network reduces the size of the model from 236MB to 37MB,a reduction of 6.378 times,and the inference speed is increased by 2 times Much(3)This paper further proposes to use the GIOU-based method to improve the network's loss function,so that the model's detection and evaluation standards during training and inference remain consistent.The improved YOLOv3-MobileNetv2 is trained with a new loss function,and its detection accuracy has been further improved.This algorithm is also trained on the data sets PASCAL VOC and MS COCO,and its experimental results verify that the improved thinking based on GIOU can effectively improve the detection accuracy of the networkThe innovation of this article is to propose a lightweight YOLOv3-MobileNetv2 target detection algorithm,extract feature maps through a lightweight network,and merge multi-scale feature information,the improved network speed has been greatly improved,while greatly reduced The storage size of the model is improved,and then GIOU is used to improve its loss function.Finally,the performance of the improved target detection system is further improved.