SVDocNet: Spatially Variant U-Net for Blind Document Deblurring

Blind document deblurring is a fundamental task in the field of document processing and restoration, having wide enhancement applications in optical character recognition systems, forensics, etc. Since this problem is highly ill-posed, supervised and unsupervised learning methods are well suited for this application. Using various techniques, extensive work has been done on natural-scene deblurring. However, these extracted features are not suitable for document images. We present SVDocNet, an end-to-end trainable U-Net based spatial recurrent neural network (RNN) for blind document deblurring where the weights of the RNNs are determined by different convolutional neural networks (CNNs). This network achieves state of the art performance in terms of both quantitative measures and qualitative results.