Lightweight detection method based on improved YOLOv4

WenLong Ruan, Yunqing LiuICCV 2022 · DOI

Abstract

In view of the large amount of model parameters and large amount of calculation when YOLOv4 target detection model is applied to unmanned vehicles, it is difficult to be applied to embedded platform or mobile terminal. In this paper, an improved YOLOv4 model was proposed, which used the MobileNetv1 as the backbone network of YOLOv4. The depth separable convolution was used to replace the ordinary convolution in the enhanced feature extraction network structure PANET of YOLOv4. By introducing the ECA attention mechanism and using DIOU_NMS post-processing strategy, the detection accuracy of the model was improved. The improved YOLOv4 model was tested on the PASCAL VOC data set. The results showed that the number of parameters was one-fifth of the original, the model size was reduced by 193M, the mAP was reduced by 0.56%, and the fps was increased by 20.4. It can meet the requirements of real-time detection of unmanned vehicles.
BibTeX
@inproceedings{ruan2022lightweight,
  author = {WenLong Ruan and Yunqing Liu},
  title = {{Lightweight detection method based on improved YOLOv4}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2022},
  doi = {10.1109/cvidliccea56201.2022.9824440},
}

Loading related content...

Figure