Gated image-adaptive network for driving-scene object detection under nighttime conditions

When applied to nighttime driving scenarios, object detection models often experience significant performance degradation due to insufficient illumination. Existing mainstream nighttime object detection techniques either cascade multiple image processing modules in an arbitrary sequence or use Non-Maximum Suppression (NMS) to eliminate redundant candidate boxes, potentially underutilizing the capabilities of the image processing modules and failing to fully exploit the relationships between candidate boxes. In this paper, we propose a Gated Image-Adaptive Network (GIANet) for low-light image enhancement and bounding box clustering, based on a novel gated learnable image signal processing with confidence information propagation clustering. Specifically, we propose a Gated Learnable Image Signal Processing (GLISP) module that learns to enhance images through object detection loss. This is achieved by learning the parameters of a series of image processing units that operate concurrently, with the outputs combined using weights learned through a gating mechanism. To better utilize the relationships among candidate bounding boxes, a confidence information passing clustering approach is adopted to enhance the correctly labeled candidate while suppressing the redundant ones. Furthermore, we develop a Spatial Transformer Feature Processing (STFP) module which constructs contextual long-range dependencies to improve GIANet’s feature extraction and representation capabilities. We evaluate GIANet using the ExDark datasets to validate the effectiveness of our method. The results demonstrate a significant performance improvement over previous state-of-the-art methods, achieving a mean Average Precision (mAP) of 50.75.