Boosting Object Detection in Extreme Low-Light Using Image Enhancement and Bounding Box Fusion

This paper presents a novel framework for evaluating and enhancing object detection performance in low-light conditions through a comparative analysis approach. We integrate the Zero-DCE++ enhancement algorithm with YOLO (You Only Look Once) object detection to create a unified system that processes both original and enhanced video streams in real-time. The framework employs a dual-stream architecture that simultaneously analyzes object detection performance on original low-light frames and their enhanced counterparts, utilizing an innovative detection merging algorithm based on Intersection over Union (IoU) metrics. Our system processes video streams at native frame rates while maintaining detection accuracy. Experimental results demonstrate significant improvements in object detection performance, particularly in challenging low-light scenarios where traditional detection methods often fail. The framework achieved an average improvement of 29.9 % in F1Score for scenes below 15 Mean Brightness threshold. The proposed system's ability to merge detection results from both streams enables robust object detection across varying lighting conditions, making it particularly suitable for real-world applications in surveillance, autonomous navigation, and intelligent transportation systems. We validate our approach through extensive testing across diverse low-light scenarios, demonstrating its effectiveness in enhancing detection reliability while minimizing false positives and computational overhead.