Lightweight object detection method based on improved YOLOv4
An improved YOLOv4 object detection algorithm is designed to solve the problems of large number of detection model parameters and difficulty in deploying on embedded devices.A lightweight Mobilenetv1 was used to replace the CSPDarketnet53 backbone feature extraction network and a deeply separable convolution was used to replace the 3×3 convolution in the subsequent network,both aiming to drastically cut down on the number of participants.An attention module named NAM was added to the detection heads to enhance the network's ability to extract detailed information.SDIoU Loss was used as the bounding box regression loss function,which can accelerate the convergence speed and improve accuracy of the detection.Experiments show that compared with YOLOv4-CSPDarknet53,the model size trained by the improved algorithm on the PASCAL VOC07+12 dataset is 47.19 M,which is about one-fifth of the original,and the FPS is increased by 40(f/s)and the mAP is increased by 2.4%.Compared with other object detection algorithms such as YOLOv4 Tiny YOLOv5s and YOLOv7,it has the characteristics of taking into account the detection speed and accuracy.
loss functionYOLOv4attention mechanismsobject detectionlightweight network
国家科技期刊平台
NSTL
万方数据
维普
