HEYOLO: Hierarchical Enhancement Network for Object Detection under Low-Light Condition

Shaohua Miao, Kai Liu, Yuhang Shi, Tangwen Yang, Ye Li, Sitao Zhang, Lingyun Zhang, Xin Wang, Xin Yu

Abstract: Current object detection models have achieved good performance on numerous benchmark datasets, but objects detection in low-light condition is a great challenge. To this end, a hierarchical enhancement network is proposed in this paper based on YOLOv5 framework, named HEYOLO. First, the network down samples input images to construct the multi-scale representation. Specifically, a context processing module (CPM), which includes the multi-rated dilated depth-wise convolutions, is used to enhance the context information of images, and simultaneously a channel enhancing module (CEM) is used to improve inter-channel information. HEYOLO adopts the end-to-end training approach and only uses normal detection loss to simplify the training process. We conduct experiments on the public dataset and the image dataset collected by our lab, to demonstrate the effectiveness of our models. The results show that HEYOLO achieves superior results. The object detection precision of the proposed framework reaches 73.2% in mAP50 on ExDark dataset and 84.3% in mAP50 on the image dataset of our lab, respectively.