Efficient document-image super-resolution using convolutional neural network

Abstract

Experiments performed by us using optical character recognizers (OCRs) show that the character level accuracy of the OCR reduces significantly with decrease in the spatial resolution of document images. There are real life scenarios, where high-resolution (HR) images are not available, where it is desirable to enhance the resolution of the low-resolution (LR) document image. In this paper, our objective is to construct a HR image, given a single LR binary image. The works reported in the literature mostly deal with superresolution of natural images, whereas we try to overcome the spatial resolution problem in document images. We have trained and obtained a novel convolutional model based on neural networks, which achieves significant improvement in terms of the peak-signal-to-noise ratio (PSNR) of the reconstructed HR images. Using parametric rectified linear units, mean PSNR improvements of 2.32, 4.38, 6.43 and 8.92 dB have been achieved over those of LR input images of 50, 75, 100 and 150 dots per inch (dpi) resolution and average word level accuracy of almost 43%, 45% and 57% on 75 dpi Tamil, English and Kannada images, respectively.

Binary Document Image Super Resolution for Improved Readability and OCR Performance

“…We are addressing the problem of SISR for a task distinctly different from the above papers: to enhance the quality of the input binary document images so that the generated images have better readability and OCR accuracy. Our work builds on what has been reported in [46] [47], which addresses the above mentioned problem for the first time and is able to obtain good PSNR, OCR character and word level accuracies starting from a downsampled version of the document image (gray in nature), the results being similar to that of the corresponding ground truth image. In the current work, we have addressed the realistic problem of obtaining an upscaled version of a binary document image, actually scanned at a low resolution.…”

An end-to-end Optical Character Recognition approach for ultra-low-resolution printed text images

“…A second popular approach has been to use super-resolution (for example, [7], [1], [8], [9] and [10]). This approach was also perhaps partially spurred on by the ICDAR 2015 Competition on Text Image Super-Resolution ( [11], [12] and [13]), which aimed to recognise text from frames extracted from video streams.…”

“…There has been considerable success with this method. For example, Pandey et al [9] reported OCR results for 75 dpi English text as 75.1% CLA (character level accuracy) and 48% WLA (word level accuracy). They achieved significantly better results on 75 dpi Tamil (90.8% CLA and 54.3% WLA) and Kannada (95.16% WLA and 70.80% WLA).…”

“…None of these researchers reported attempting OCR with 60 dpi images: whilst even 50 dpi images were improved using super-resolution (for example, Pandey and Ramakrishnan [9] report an average PSNR of 16.89 for super-resolved 50 dpi images and 19.36 for 75 dpi images), the OCR systems are relatively poor at extracting text from such noisy images, and so the accuracy rate would be even lower for super-resolved 60 dpi images. We also note at this point that 60 dpi images contain only about 60% as many pixels as the corresponding 75 dpi images, making the task significantly harder.…”

MSCE: An Edge-Preserving Robust Loss Function for Improving Super-Resolution Algorithms

“…Since the advent of deep learning technologies in the past decade, superresolution algorithms have shown remarkable improvement in the quality of the reconstructed image. Most of the work reported in the literature have used mean square error (MSE) loss function to minimize the error between the reconstructed model output and the ground truth image [2] [3] [4] [5] [6] [7]. Minimizing this loss function may reduce the high frequency content in the reconstructed image and thus may blur the edges in it.…”