A New Modeling Approach for Extremely Low Light Video Using Video Enhancement Techniques

Abstract

Enhancing low light videos has been quite a challenge over the years. Low dynamic range and significant noise are always concerns with video captured in low light. This is a thesis towards PhD degree makes a contribution to the subject of video enhancement in low light. Three models are proposed for extremely low light low quality video enhancement. newlineTo increase the quality of videos in Model 1, the noise in the videos must first be decreased before any enhancing can be attempted. This method uses a Kalman filter to reduce the amount of noise in extremely low-light video. A Nonlocal Means Filter (NLM) can assist remove the remaining noise after amplification, while this technique cannot eliminate all noise. For amplification, the Gray World approach is applied. After decreasing noise with the NLM filter, image enhancement of extremely low light video is achieved. newlineIn Model 2, The often faced challenge to preserve naturalness of digital videos arise when they are taken under non uniform illumination either due to poor light environment or due to uneven lighting on the object surface or due to prevalence of fixed noise patterns. The first step recommended in the study is to decompose the frame into reflectance and illumination. The Bright Pass Filter (BPF) is used for this purpose. Next step is to obtain uniform illumination. Towards this apply the technique called Adaptive Gamma Correction with Weighted distribution of illumination. After this combine reflectance and uniform illumination. In order to enhance further, The Dual combination of background Subtraction and Colour Consistency Enhancement (DSCE) is recommended. This considerably reduces fixed pattern noise enhancing the quality of video frames. newlineIn Model 3, deep learning based system is adopted. For this purpose a Generative Adversarial Network (GAN) based Extremely Dark Video Enhancement Network (GEVE) model is proposed. The GEVE model consisting of two networks namely generating network and discriminating network. The generating continuously generates data