Research On Real-time Object Detection Based On Lightweight CN

With the wide application of deep learning in image processing,the target detection algorithm based on convolutional neural network gradually replaces the traditional algorithm.However,the target detection network has the problems of large number of parameters,high computational complexity and slow detection speed.It usually needs to be deployed on equipment with high computing ability.It is difficult to meet the real-time requirements in small equipment with limited computing resources,which undoubtedly increases the cost of equipment.To solve this problem,this paper proposes an improved lightweight target detection network based on YOLOv4-tiny algorithm,and further improves its detection accuracy and speed.YOLOv4-tiny has fewer parameters than YOLOv4,but the detection accuracy is low,and the demand for computing power of deployment equipment is still high.Aiming at the problem that the main feature extraction network structure of YOLOv4-tiny is simple and it is difficult to extract the feature information effectively,this paper proposes to replace the CSP structure in the main feature extraction network by Bneck＿E structure,which effectively reduces the number of parameters of the main feature extraction network while increasing the network depth.Aiming at the problem of low detection accuracy of small targets and high missed detection rate in complex scenes,a lightweight attention mechanism ECA is added in the classification regression network to improve the accuracy of the model while introducing fewer parameters.Two-way feature fusion is adopted,and depthwise separable convolution is used to sample the shallow features,which improves the detection accuracy of small targets and reduces the number of parameters.Experiments on PASCAL VOC data set show that the improved algorithm m AP increases by 4.4%,FPS increases by 15.7%,and the number of parameters is only 36% of YOLOv4-tiny.In order to further improve the detection speed of the model on small embedded devices,this paper adopts the method of cutting redundant channels based on the gamma factor of BN layer,and prunes channels on some network layers.At the pruning rate of 25%,the m AP of the model is reduced by only 0.59%,the number of parameters is reduced by 19%,and the FPS is increased by 8.8%,which further improves the detection speed of the model.In this paper,Raspberry Pi 4B is selected as the embedded device,NCNN forward reasoning framework is selected during deployment,and the model is quantified by Int8.The experimental results show that the proposed lightweight target detection network only needs 173 ms to detect a single image on Raspberry Pi 4B,and the detection time is reduced by 43%.It has good real-time performance,and the missing rate is lower than that of YOLOv4-tiny,which proves the effectiveness and feasibility of the lightweight model in this paper.