Image text deblurring by convolutional neural networks

Image deblurring is one of the most studied topics by researchers in recent years, due to the frequent use of cameras in various fields such as surveillance, photography, etc. One of the areas in which the cameras available in smartphones are used is to take images of documents in order to save or send them over the Internet. Image degradation is inevitable during the acquisition or transmission process. With the advancements in computational techniques and tools, image degradation can be avoided or corrected to a great level, however, the quality of the acquired images is still insufficient for many applications. This calls for the development of the most advanced digital photo recovery tools. This thesis focuses on creating an architecture using deep learning methods for blind recovery of blurred text images. Autoencoder architecture that is based on U-net, Resnet, and grouped dilated convolution for deconvolution is proposed. The U-net architecture consists of two parts, an encoder for feature extraction and a decoder for image regeneration. Resnet and Dilated convolution architecture without reducing the size of the image is used. The model is capable of producing realistic latent images with photo-quality effects in a reliable manner. Extensive testing on simulated and realistic blurry images showed that the proposed network is comparable to current methods. Compared with traditional general-purpose de-interference methods, the proposed deblurring algorithm can produce more satisfactory results on text images. The encouraging experimental results demonstrate the effectiveness of the proposed method.