DFE-Net: detail feature extraction network for small object detection

The universal object detector has the problem of low detection accuracy and high missed detection rate when detecting small objects. In this paper, we propose a novel deep learning model called detail feature extraction network (DFE-Net) based on YOLOv5. Firstly, we remove C5 layer from the baseline backbone and introduce the detail feature enhancement module. In this module, we incorporate extended separable convolution to enable the model to effectively capture fine-grained information about small objects. Secondly, we propose a novel decoupled head based on attention. The mixed attention module assigns different weights to the classification and regression tasks in separate branches, effectively improving detection performance. Finally, the SIoU loss function is introduced to speed up the convergence of the model. The experimental results indicate that our proposed DFE-Net, when compared to the baseline, not only exhibits a significantly higher improvement in accuracy on the general small object dataset, but also reduces the parameter count by 57.7%. The mAP \(_{50}\) ${}_{50}$ and mAP \(_{50-95}\) ${}_{50-95}$ improve by 3.7% and 2.1% on the VisDrone2019 dataset, and by 5.9% and 5.1% on the TT100K dataset.