Bio-Inspired Low-Light Image Enhancement with Large Kernel Convolution and Attention

Abstract

Nighttime driving safety remains a critical challenge in modern transportation systems: insufficient ambient lighting significantly degrades visual perception quality, adversely affecting both human drivers and advanced driver-assistance systems (ADAS) and directly threatening road users’ safety. Traditional image enhancement methods often suffer from color distortion and visual artifacts, whereas existing deep learning approaches typically require paired training data and incur substantial computational overhead. To address these limitations, this paper presents BLEN (bio-inspired low-light enhancement network), a zero-reference deep learning framework that integrates biological vision principles with efficient convolutional architectures. Specifically, BLEN leverages Retinex theory for illumination–reflectance decomposition, is inspired by and functionally approximates lateral inhibition mechanisms for edge enhancement, and incorporates a Large-Kernel Convolution with Attention (LKCA) module that reduces the parameter count of the LKCA encoder block by 76% (0.56 M vs. 2.34 M for a standard 13 × 13 convolution) relative to standard large-kernel operations. Extensive experiments on the SICE and LOL benchmarks demonstrate that BLEN achieves state-of-the-art performance among real-time, edge-deployable zero-reference methods on the SICE benchmark, yielding a peak signal-to-noise ratio (PSNR) of 23.67 ± 0.14 dB and a structural similarity index measure (SSIM) of 0.891 ± 0.004 on SICE while maintaining 2.10 M parameters (2.1 MB in INT8, 8.4 MB in FP32). Furthermore, the proposed method enables real-time inference at 31 frames per second (FPS) on embedded platforms, including the HiSilicon SS928 and Jetson Nano, demonstrating that the proposed method is an efficient and effective front-end for camera-based ADAS perception on automotive-grade edge hardware.