Mutual Support and Promotion: Learning Structure Compensation and Context Completion for Low-Light Vision

In recent years, visual perception tasks (e.g., object detection) in harsh scenes (e.g., low-light scenes) have received extensive attention due to their importance in various practical applications (e.g., autonomous driving). However, when addressing the above tasks, existing works are often too simplistic and straightforward to characterize different types of features, and often only focus on a single task while neglecting the performance of multiple tasks. To break the limitations of existing methods, this paper proposes a Structure Compensation and Context Completion Network (SC
³
N) for solving perception tasks in low-light scenes. Specifically, we first construct an object detection network and a self-supervised low-light image enhancement network to jointly extract various task-related features while meeting the task requirements. Furthermore, in order to reasonably utilize the effective information held by complex and different kinds of features, we design a structure compensation module and a context completion module to realize the information exchange between different tasks. Through the proposed bidirectional information interaction framework, the method could realize the mutual transfer and dynamic guidance of feature-level structural information and feature-level context information, thereby enabling the network to jointly improve low-light performance low-light images enhancement and detection. Finally, extensive comparative experiments fully verify the superiority of the proposed method compared with other state-of-the-art methods in both enhancement and detection. In addition, we also design a series of ablation analyses for each module to verify the effectiveness of the proposed method.