Photonic-inspired multiscale frequency fusion for low-light YOLO detection

Abstract

Low-light object detection remains challenging because insufficient illumination, weak local contrast, noise, color distortion, and blurred object boundaries can degrade the visual information required by object detectors. Although image enhancement is commonly applied before detection, a visually brighter image does not necessarily yield better detection performance. This paper presents a photonic-inspired multi-scale frequency fusion enhancement operator, termed PIISEMF, for detection-oriented low-light YOLO processing. PIISE-MF operates on the LAB luminance channel and combines multi-scale Gaussian filtering, a darkness-adaptive weight, local gamma-based illumination compensation, and bounded band-pass residual fusion. A class-balanced and luminance-stratified 50-image ExDark development set was used to evaluate the original image, fixed conventional enhancement branches, and the proposed PIISE-MF weak and strong branches. Official ExDark annotations were used to calculate true positives, false positives, false negatives, Precision, Recall, and F1-score using class-aware IoU-based matching. On this development set, the PIISE-MF weak branch achieved 83 true positives, 46 false positives, and 59 false negatives, improving F1-score from 0.5912 for the original branch to 0.6125. This result is also slightly higher than the evaluated CLAHE branch (0.6081). The current implementation is conducted in the digital domain for algorithmic validation. Its core linear operations are deliberately designed to be compatible with photonic analog computing architectures, providing a feasible pathway for future hardware acceleration in edge-side vision systems. Detection-guided adaptive branch selection remains under development and is not reported as a completed result in this manuscript.