Improved YOLOv4 Based on Dilated Convolution and Focal Loss
Feng Dan Dong, Mangui Liang, Guangchao Wang · 2021
With the increasing application of object detection algorithms in practice, the single-stage detector YOLOv4 excellent in both detection precision and inference speed has been widely used in various scenarios. In order to further improve the small object detectability of YOLOv4, this paper proposes to combine the 8-times downsampling feature map of YOLOv4 with the feature map that is outputted by the second residual block of CSPDarknet-53 and processed via the Hybrid Dilated Convolution (HDC), to obtain new detection features. In addition, this paper comes up with an idea of using Focal Loss to improve the negative sample confidence formula for the loss function and alleviate the proportion imbalance between positive and negative samples of YOLOv4. The experimental results show that for the specific test set where small objects account for 47.7% of the total, the average precision and recall of the improved YOLOv4 increase by 8.8% and 16%, respectively, compared with those of the original YOLOv4. Similarly, for the PASCAL VOC test set, the average detection precision of the improved YOLOv4 is 3.4% higher than that of the original YOLOv4.