LiteLTC‐Net: Image Enhancement via Liquid Dynamic Features and Lightweight Attention

Abstract

ABSTRACT Low‐light image enhancement is essential for real‐world applications like surveillance and autonomous driving, but existing deep models are often too heavy for edge deployment. To address these limitations, this paper introduces LiteLTC‐Net, a compact neural network for low‐light image enhancement that leverages a simplified liquid time‐constant (LiteLTC) layer to enable adaptive feature transformation guided by local illumination. LiteLTC‐Net integrates the dynamic modelling capacity of liquid neural networks. This allows the model to effectively enhance image details under varying lighting conditions. The network incorporates a LiteHybridAttention mechanism, which combines lightweight channel and spatial attention, and a lite local enhancement module for contrast refinement. Multi‐scale features are fused via residual connections and upsampling, culminating in a lightweight convolutional decoder. With only 9.9K parameters, LiteLTC‐Net achieves strong performance on the MIT‐Adobe FiveK dataset (LPIPS = 0.046, SSIM = 0.901), highlighting the effectiveness of liquid neural dynamics in ultra‐lightweight image enhancement models.