ENHANCING REAL-TIME OBJECT DETECTION IN LOW-LIGHT CONDITIONS USING ZERO-DCE AND SUPER-RESOLUTION GANs: A YOLO-BASED APPROACH

Abstract

Low-light conditions significantly degrade the performance of real-time object detection systems. This study proposes a novel pipeline that integrates Zero-Reference Deep Curve Estimation (Zero-DCE), which has been used to enhance the low-light image, and Enhanced Super-Resolution Generative Adversarial Networks (ESRGANs) for improving the object detection accuracy in poor illumination condition for resolution refinement. The enhanced images are then processed through a YOLO-based detector for real-time object identification. Zero-DCE is leveraged to enhance image illumination without requiring reference images or paired datasets, ensuring efficient and adaptive enhancement across diverse lighting conditions. Following enhancement, ESRGAN is applied to increase the perceptual quality and fine-grained details of the images, enabling the detection model to capture subtle features that are often lost in low-light inputs. This dual stage preprocessing significantly improves the visibility and quality of the input images, directly benefiting object detection performance. The experimental evaluation, conducted on low-light datasets, demonstrates substantial improvements in detection accuracy, precision, and recall metrics. Furthermore, the proposed pipeline maintains real-time performance that can be suitable for surveillance, autonomous navigation, and security applications.