Deep Learning Based Exposure Correction for Image Exposure Correction with Application in Computer Vision for Robotics

This paper presents a convolutional neural network that is able to minimize the effects of poorly exposed digital images in post processing, outperforming state-of-the-art networks on two different conditions: images that are too dark, and images where the irradiance exceeds both the lower and upper bounds of the sensor. The scene's illumination, camera aperture, exposure time, and ISO sensitivity present a direct impact on the quality of the resulting image. Optimize these parameters, however, is a challenging task, which often leads to images that are either too bright, too dark or out of focus. In order to restore over-exposed and under-exposed images, we propose a Generative Adversarial Network (GAN) which learns how to restore an image, preserving and enhancing the contents of the input image, as well as complete regions where data was completely lost due to clipping. Our results show the proposed network is able to enhance color, fulfill regions with missing data, and preserve structural details.