DeepDeblur : text image recovery from blur to sharp

• Main Authors: Mei, Jianhan, Wu, Ziming, Chen, Xiang, Qiao, Yu, Ding, Henghui, Jiang, Xudong
• Other Authors: School of Electrical and Electronic Engineering
• Format: Article
• Language: English
• Published: 2021
• Subjects: Engineering::Electrical and electronic engineering; Text Deblurring; Convolutional Neural Network
• Online Access: https://hdl.handle.net/10356/151740
• Institution: Nanyang Technological University

Digital images could be degraded by a variety of blur during the image acquisition (i.e. relative motion of cameras, electronic noise, capturing defocus, and so on). Blurring images can be computationally modeled as the result of a convolution process with the corresponding blur kernel and thus, image deblurring can be regarded as a deconvolution operation. In this paper, we explore to deblur images by approximating blind deconvolutions using a deep neural network. Different deep neural network structures are investigated to evaluate their deblurring capabilities, which contributes to the optimal design of a network architecture. It is found that shallow and narrow networks are not capable of handling complex motion blur. We thus, present a deep network with 20 layers to cope with text image blur. In addition, a novel network structure with Sequential Highway Connections (SHC) is leveraged to gain superior convergence. The experiment results demonstrate the state-of-the-art performance of the proposed framework with the higher visual quality of the delurred images.