Development of Deep Learning Techniques for Image Super Resolution

The images to be used in many of the real-life applications, such as medical imaging, intelligent transportation systems and space explorations, are not of a sufficient quality in view of the degradation processes associated with the image capturing devices. In recent years, deep neural networks have emerged as a sophisticated tool for image restoration. However, many of the existing deep neural networks for image restoration employ a large number of parameters for providing high performance, thus prohibiting their deployment in applications with the constraints on memory and power consumption. Hence, the design of high-performance image restoration convolutional neural networks that employ small number of parameters is of paramount importance. As the performance of deep networks is closely related to the richness of features produced by them, the objective of the thesis is to design deep image restoration neural networks that are capable of producing rich sets of features by using only a small number of parameters. In this thesis, this objective is met by using the suitable prior information associated the degradation processes of the image capturing devices. Three specific degradation models, namely, bicubic downsampling, Gaussian blurring coupled with downsampling and JPEG compression blocking, are considered in designing a number of deep light-weight image restoration networks. With regard to the first degradation model, i.e., bicubic downsampling operation, several image super resolution networks using the different prior information about this operation are developed. Specifically, four different prior information, namely, multi-scale feature generation, guided feature generation, efficient feature fusion and sparsity prior, are used for developing light-weight image super resolution networks. As to the second degradation model, i.e., Gaussian blurring coupled with downsampling, two deep networks, in which the blurred version of the high-quality images are used in the context of global residual learning as the prior information, are proposed. Finally, with respect to the third degradation model, i.e., JPEG blocking artifacts, two information, namely, robust features generated by the maxout activation units and the high-frequency components generated by the fractal neural networks, are used as prior information to propose couple of image restoration networks. Extensive experiments are carried out to validate the effectiveness of the various ideas and schemes developed in this thesis for improving the quality of degraded images