Phydiisp: a physics-guided differentiable pipeline for low-light machine vision

Jingchao Hou, Garas Gendy, Guanghui He

Abstract: In low-light environments, machine vision tasks often suffer from performance degradation because traditional Image Signal Processing (ISP) pipelines are primarily optimized for image quality metrics such as Peak Signal-to-Noise Ratio and Structural Similarity Index, which do not adequately address the specific needs of these applications. Existing methods fall short in enhancing the critical image features required for computer vision tasks under challenging lighting conditions. To address this, we introduce PhyDiISP, a physics-guided, differentiable ISP pipeline designed to improve machine vision performance in low-light scenarios. PhyDiISP integrates traditional ISP design principles with physical insights, including demosaicing for RAW-to-RGB conversion, global tone mapping to adjust overall brightness, and Multiscale Retinex-based enhancement to tackle low-light challenges. Experimental results show that PhyDiISP outperforms existing ISP methods in object detection accuracy across standard benchmarks by effectively enhancing key image features. Furthermore, when trained with L1 loss and aligned with ground truth on datasets of dark-light environments and real RAW-to-RGB conversions, it demonstrates competitive image quality. These results confirm that PhyDiISP is a viable and effective solution for real-world low-light machine vision applications.