DEEP LEARNING FOR DEFOCUS DEBLURRING

Abstract

In this dissertation, several techniques are developed to facilitate the application of deep learning in the recovery of defocused images. The first is a data generation pipeline to help network training without collecting any real paired training samples. We proposed an efficient approach for synthesizing blurred/sharp image pairs which enable effective training of a non-uniform deblurring model with sufficient statistical characteristics, without attempting to simulate real-world images. The second is the introduction of an unsharp mark filtering technique which enables a better restoration in regions that contain dense image edges(the most difficult regions for recovery). The last one is the introduction of attention mechanism for handling spatially varying blurring. Such an attention mechanism enables the network to have an adaptive processing over different image regions with different blurring effects.