CNN Based Image Restoration

This work proposes an artificial neural network model to restore images damaged by inadequate sensor exposure, saturation, and underexposure, at the time of acquisition. The problem has significant relevance in computational and robotics vision applications, especially when obtaining images of scenes with non-Lambertian surfaces, as well as natural images where the sensor limitation or optical arrangement prevents the scene details from being adequately represented in the captured image. We chose to model an alternative based on deep neural networks, which is adequate, considering the variability in equipment and photography techniques, along with several uncontrolled variables affecting the process. Given a set of synthetic and real image pairs, the representation structure is able to converge into a robust image enhancement model. The proposal incorporates recent advances made by convolutional networks on issues such as semantic segmentation and classification in images. The development and evaluation of the research results are primarily quantitative, using qualitative analysis when appropriate. Results measured by different indicators of image quality indicate that the proposed neural network model can improve images damaged by an amount of 3% on the best scenario on the PSNR metric.