Lightweight Zero-DCE-Driven YOLO for Nighttime Low-Light Object Detection

To address the issue of difficulty in balancing accuracy improvement and computational overhead in the joint deployment of low-light enhancement and target detection, this paper constructs a lightweight Zero-DCE-driven YOLO night low-light target detection framework. Depthwise separable convolution, spatial attention, and SE attention are all common components. This paper does not consider them as independent structural innovations. Instead, it analyzes their applicable conditions based on the brightness distribution, local gradient, and channel response characteristics of low-light images, and optimizes the enhancement and detection jointly to enable them to work collaboratively for detecting targets. ExDark experiments show that the complete model improves mAP@0.5 from 74.6% to 87.2%, corresponding to an improvement of 12.6 percentage points. The end-to-end inference speed reaches 41.3 FPS, compared with 45.2 FPS for the original YOLOv8s. Therefore, the proposed method achieves a substantial accuracy improvement at the cost of a limited 3.9 FPS reduction while maintaining real-time inference capability.