Research on Vehicle Detection Based on Improved YOLOv5s
[Purposes]Aiming at the problems of complex model and low detection accuracy of the cur-rent object detection algorithm in the field of vehicle detection,a vehicle detection research based on im-proved YOLOv5s is carried out.[Methods]The Ghost module was replaced with the original YOLOv5s backbone network to achieve the purpose of model pruning,which reduced the complexity of the im-proved network model and solved the problem of large network model;Then the Squeeze and Excitation attention mechanism is introduced to extract more important feature information to improve detection ac-curacy.The data sets used in this study are all images of cars,and on the vehicle detection dataset,a to-tal of 12 786 pictures,the dataset is divided into 8∶1∶1.And among them,the training set is 10 228 pic-tures,the test set and verification set are 1 279 pictures and the method of comparative experiment was used in this study.[Findings]Experimental results show that compared with the original YOLOv5s,the average accuracy of the improved network model is increased by 3%,the accuracy and recall rate are in-creased by 1.9%and 3.2%,respectively,and the model size is reduced by 42%.[Conclusions]The im-proved network model effectively reduces the complexity of the model,saves costs and improves the de-tection accuracy.
deep learningobject detectionattention mechanismYOLOv5s
NETL
NSTL
万方数据
