Reparameterizing Residual Unit for Real-time Maritime Low-light image Enhancement

Abstract

Video surveillance is critical in the maritime industry. However, the inescapable low-light situation places a limitation on video surveillance advancement. At the same time, the high precision of deep learning brings high computational and memory requirements to its training and inference stages. However, high precision and high resource consumption are the characteristics of deep learning. To more effectively deploy the learning-based low-light enhancement method on the terminal device, we adopted the reparameterization technology in the enhancer model to reduce the number of additional calculations (named RepMConv). Specifically, we use linear combinations of inconsistent kernel sizes in the training phase and fold them back to normal convolutions in the inference phase. Convolution kernels with different sizes can effectively extract enhancer’s significant edge and texture information by providing different receptive fields. We first embed RepMConv into the residual block to improve the learning efficiency of the residual block. Finally, we complete our enhancer network in a multi-scale structure of encoder-decoder. Experimental results show that our proposed Enhancer can achieve high-quality maritime low-light image enhancement while maintaining breakneck inference speed.