Deep over and Under Exposed Region Detection

Abstract

The camera sensors often fail to capture all the brightness intensities present in the visible spectrum of light. This is due to the limited dynamic range of the sensor elements. When bright light falls on a camera sensor, it is not appropriately measured. The recorded brightness values that fall outside the sensor’s dynamic range are stored as the minimum or maximum value depending on the bit-depth of the sensor. This results in a loss of information and undesirable artifacts in the form of blown-out areas, referred to as over- and under-exposed regions. In this study, we propose to detect these areas in an image using deep learning tools. Our approach uses semantic segmentation to mark the under, over, and correctly exposed regions in the image. We have created a new dataset containing 4928 images to train and test the performance of the model using a pre-trained state-of-the-art model architecture and re-trained it on our custom dataset. To the best of our knowledge, this is the first attempt to use semantic segmentation and transfer learning methods to identify these regions in an end-to-end fashion. We obtain a Dice score and a Jaccard score of 0.93 and 0.86, respectively, which are better than the state-of-the-art methods. The quantitative and qualitative results show that the proposed method outperforms several existing methods for identifying the over and the under-exposed regions. We will make the dataset public for research work.