ShabbyPages: A Reproducible Document Denoising and Binarization Dataset

Alexander GroleauKok Wei CheeStefan LarsonSamay MainiJonathan Boarman

Mar 2023

摘要

Document denoising and binarization are fundamental problems in the documentprocessing space, but current datasets are often too small and lack sufficientcomplexity to effectively train and benchmark modern data-driven machinelearning models. To fill this gap, we introduce ShabbyPages, a new documentimage dataset designed for training and benchmarking document denoisers andbinarizers. ShabbyPages contains over 6,000 clean "born digital" images withsynthetically-noised counterparts ("shabby pages") that were augmented usingthe Augraphy document augmentation tool to appear as if they have been printedand faxed, photocopied, or otherwise altered through physical processes. Inthis paper, we discuss the creation process of ShabbyPages and demonstrate theutility of ShabbyPages by training convolutional denoisers which remove realnoise features with a high degree of human-perceptible fidelity, establishingbaseline performance for a new ShabbyPages benchmark.

展开全部

图表提取