Enhancing visibility of low-light images using deep learning techniques

Abstract

Visualizing an image in the low light is still a challenging and an unaccomplished objective, due to low Signal Noise Ratio (SNR) and low photon count. Though many techniques on image processing have been proposed, such as deblurring and denoising, to increase the visibility of the image in the darkness, they have certain drawbacks and limitations. The model proposed in this paper is deep learning pipeline. We have trained two models in order to enhance the image, one is based up on the convolutional network with raw short exposure image with reference of its corresponding long exposure image. The second model is based on the separation of an image into its RGB (Red, Blue, Green) channels, and training an individual model for each channel. Both the models are tested and promising results are obtained in terms of the SNR, on the new datasets.