DyHead-YOLOv5 Based on Improved Object Detection Heads with Attentions
Xingya Yan, Fengfeng Liu · 2023
A survey of object detection algorithms has revealed that the low accuracy of most lightweight object detection algorithms,More and more people are working on various algorithms. In this paper, the model training is carried out by improving the YOLOv5 object detection algorithm, adding a Dynamic Head framework based on attention mechanism to the original Head layer, modifying the default normalization method Group Normalization (GN) to Batch Normalizatio (BN) in the original framework, and then conducting experimental research on the VisDrone datasets and VOC2012 datasets. The results show that compared with the YOLOv5s model, key indicators such as mean of Average Precision (mAP)@0.5 and [email protected]:0.95 are improved, which can achieve better recognition effect on the dataset without significantly increasing the complexity of the model.