A Novel Image Colorization Architecture and UnderWater Image Color Correction Using Deep Neural Networks

Abstract

Image colorization has been an interesting and exciting research topic due to its “multi-modal” nature. Before the onset of deep learning, image colorization systems were not end-to-end and required a tremendous amount of effort to colorize a grayscale image. With the advancement of deep learning, end-to-end systems are being designed for the task. The thesis comprises of literature survey of different methods to understand the problem structure and the already prevalent solutions. The thesis presents an ensemble encoder-decoder approach to tackle the problem. Transfer learning has been used to leverage the power of deep learning. Pretrained networks ResNet50 and DenseNet121 have been used to extract multilevel distinct and varied features. The features are then fused. Different feature fusion techniques have been explored. The fused features are then propagated to the decoder module along with encoder features as skip connections. The work is extended to underwater image color correction. A very similar setup has been used for the underwater image color correction problem. The results obtained are reasonably competitive with respect to subjective and referencebased image quality assessment metrics.