Improvement of YOLOV5 Model Based on the Structure of Multiscale Domain Adaptive Network for Crowdscape

Abstract

In this paper, we propose an improved model DAN-YOLOV5 based on YOLOV5. First, we use a mosaic enhancement strategy, which creates a large number of new samples on the existing VOC2007 dataset. Second, an innovative adaptive network module DAN is used on top of YOLOV5. The adaptive network module DAN is used to fuse features from same-layer scenes or cross-layer scenes. Finally, the experimental results show that the accuracy of the YOLOV5 dataset enhanced with shear-mixing and mosaic enhancement strategies is 71.02%, which is 13.56% better than the unenhanced data, and the average accuracy Figure is 80.05%, which is 33.11 percentage points better than the data. Applying the adaptive network module DAN to the YOLOV5 model, it improves the accuracy by 2.61% relative to YOLOV5 at 75.28%. Achieving such experimental results without increasing the computational effort and complexity at the grassroots level is well worth studying.