Coarse-To-Fine Luminance Estimation for Low-Light Image Enhancement in Maritime Video Surveillance

Abstract

Captured images in maritime video surveillance under non-uniform illumination conditions easily suffer from low contrast and details loss. The low-quality images may significantly result in negative effects in practical applications, e.g., target detection, recognition, classification and tracking, etc. Increasing attention has recently been paid to improve the quality of low-light images via computer vision techniques. In this paper, we propose to establish a two-step luminance estimation framework to enhance low-light images. In particular, the coarse luminance is firstly estimated using traditional Max-RGB which extracts the highest pixel values in each color channel. The refined luminance is obtained by introducing a weighted variational model which has the capacities of structure-preserving and texture-smoothing. Based on the estimated well-constructed luminance, the enhanced low-light images are obtained by combining Retinex model with its extended version. The image quality is further improved through a BM3D-based denoising approach. Experimental results on both synthetic and realistic low-light images have demonstrated the satisfactory imaging performance in terms of quantitative and qualitative evaluations.