Blind Source Separation Techniques for Detecting Hidden Texts and Textures in Document Images

摘要: Blind Source Separation techniques, based both on Independent Component Analysis and second order statistics, are presented compared for extracting partially hidden texts textures in document images. Barely perceivable features may occur, instance, ancient documents previously erased then re-written (palimpsests), or transparency seeping of ink from the reverse side, watermarks paper. Detecting these can be great importance to scholars historians. In our approach, is modeled as superposition a number source patterns, simplified linear mixture model introduced describing relationship between sources multispectral views itself. The problem detecting patterns that barely visible color image thus formulated one separating various mixtures. Some examples an extensive experimentation with real shown commented.