Building an Image Enhancer using Deep Learning and SICE Techniques

Most of the images captured on digital devices like cameras, mobiles are often under exposed or over exposed to light due to inappropriate lighting conditions, which is a cause of loosing detailing of the picture. To adjust the lighting in the outcome, there are various techniques like single image contrast enhancement which are trained on a single image to spotlight the defects in the image and correct it wherever needed. This could solve the irregularities to some extent, but may not give satisfactory results in all possible scenarios. Hence, we need to train a memory (algorithm in this case) on multiple images, which could memorise a defect and its corresponding resolution tactics. All of this knowledge could be used at once to identify multiple blemishes in the input and corresponding fixes could be made for each of them. For this purpose of knowledge extraction, Convolutional neural networks (CNN) are employed on the dataset which will study and identify the problems like darkness, over exposure, blurred images and apply the remedies on them. Low Light Image/ Video enhancing (LOL) dataset is used for this purpose which has 500 pairs of defective and corresponding corrected images. CNN is trained easily on the dataset to provide significantly better results over the existing SICE techniques.