Investigation Results of State of Art Algorithms in Denoising Text Images

One of the crucial issues in acquired images is the corruption due to various kinds of noise. Recovering vital information of the original image from the noisy image is the purpose of image denoising. Denoising text images is of special interest since they convey important information. They are, however, corrupted by noise due to scanning, transfer or digitization. Old document images which date back to many years are affected by noise and deteriorations more seriously. Some of the deteriorations include ink-leakage from the black page, brown depigmentation of paper, fading text, background points. Reduction of noise, or denoising, is a vital step in the image processing of documents. Various methods for decreasing noise have been proposed by researchers already. Weighted Nuclear Norm Minimization (WNNM) will employ the low-rank standard. This algorithm uses low-rank models and produces good denoising results. This thesis applies WNNM method to different text images and compares this method with some other traditional denoising methods, such as Block Matching and 3D Filtering (BM3D), Non-local Centralized Sparse Representation (NCSR), and Expected Patch Log Likelihood (EPLL). A preliminary investigation of the various methods is also carried out for some ancient document images. The results show that using the WNNM method and denoised images have a higher peak signal-to-noise ratio (PSNR) than other methods. For example, for 𝜎�� = 10, WNNM denoised images in average have PSNR=32.21, while BM3D denoised images in average have PSNR=31.84, EPLL denoised images have PSNR=31.47, and NCSR denoised images have PSNR=31.93.