Unsupervised neural domain adaptation for document image binarization

Binarization is a well-known image processing task, whose objective to separate the foreground of an from background. One many tasks for which it useful that preprocessing document images in order identify relevant information, such as text or symbols. The wide variety types, alphabets, and formats makes binarization challenging. There are multiple proposals with solve this problem, classical manually-adjusted methods, more recent approaches based on machine learning. latter techniques require large amount training data obtain good results; however, labeling portion each existing collection documents not feasible practice. This common problem supervised learning, can be addressed by using so-called Domain Adaptation (DA) techniques. These take advantage knowledge learned one domain, labeled available, apply other domains there no data. paper proposes method combines neural networks DA carry out unsupervised binarization. However, when both source target very similar, adaptation could detrimental. Our methodology, therefore, first measures similarity between innovative manner determine whether appropriate process. results reported experimentation, evaluating up 20 possible combinations among five different domains, show our proposal successfully deals new without need