Image enhancement and denoising in extreme low-light conditions

Abstract

Image noise refers to the specks of false colors or artifacts that diminish the visual quality of the captured image. It has become our daily experience that with affordable smart-phone cameras we can capture high clarity photos in a brightly illuminated scene. But using the same camera in a poorly lit environment with high ISO settings results in images that are noisy with irrelevant specks of colors. Noise removal and contrast enhancement in images have been extensively studied by researchers over the past few decades. But most of these techniques fail to perform satisfactorily if the images are captured in an extremely dark environment. In recent years, computer vision researchers have started developing neural network-based algorithms to perform automated de-noising of images captured in a low-light environment. Although these methods are reasonably successful in providing the desired de-noised image, the transformation operation tends to distort the structure of the image contents to a certain extent. We propose an improved algorithm for image enhancement and de-noising using the camera’s raw image data by employing a deep U-Net generator. The network is trained in an end-to-end manner on a large training set with suitable loss functions. To preserve the image content structures at a higher resolution compared to the existing approaches, we make use of an edge loss term in addition to PSNR loss and structural similarity loss during the training phase. Qualitative and quantitative results in terms of PSNR and SSIM values emphasize the effectiveness of our approach.