Deep Image Enhancement for Ill Light Imaging.

Rizwan Khan,You Yang,Qiong Liu,Jialie Shen,Bing Li

DOI: https://doi.org/10.1364/josaa.410316

2021-01-01

Journal of the Optical Society of America A

Abstract:Imaging in the natural scene under ill lighting conditions (e.g., low light, back-lit, over-exposed front-lit, and any combinations of them) suffers from both over- and under-exposure at the same time, whereas processing of such images often results in over- and under-enhancement. A single small image sensor can hardly provide satisfactory quality for ill lighting conditions with ordinary optical lenses in capturing devices. Challenges arise in the maintenance of a visual smoothness between those regions, while color and contrast should be well preserved. The problem has been approached by various methods, including multiple sensors and handcrafted parameters, but extant model capacity is limited to only some specific scenes (i.e., lighting conditions). Motivated by these challenges, in this paper, we propose a deep image enhancement method for color images captured under ill lighting conditions. In this method, input images are first decomposed into reflection and illumination maps with the proposed layer distribution loss net, where the illumination blindness and structure degradation problem can be subsequently solved via these two components, respectively. The hidden degradation in reflection and illumination is tuned with a knowledge-based adaptive enhancement constraint designed for ill illuminated images. The model can maintain a balance of smoothness and contribute to solving the problem of noise besides over- and under-enhancement. The local consistency in illumination is achieved via a repairing operation performed in the proposed Repair-Net. The total variation operator is optimized to acquire local consistency, and the image gradient is guided with the proposed enhancement constraint. Finally, a product of updated reflection and illumination maps reconstructs an enhanced image. Experiments are organized under both very low exposure and ill illumination conditions, where a new dataset is also proposed. Results on both experiments show that our method has superior performance in preserving structural and textural details compared to other states of the art, which suggests that our method is more practical in future visual applications.