Minimum Amount of Text Overlapping in Document Separation
Antonio Boccuto, Ivan Gerace, Valentina Giorgetti · viXra · 2018
We consider a Blind Source Separation problem. In particular we focus on reconstruction of digital documents degraded by bleed-through and show-through effects. In this case, since the mixing matrix, the source and data images are nonnegative, the solution is given by a Nonnegative Factorization. As the problem is ill-posed, further assumptions are necessary to estimate the solution. In this paper we propose an iterative algorithm in order to estimate the correct overlapping level from the verso to the recto of the involved document. Thus, the proposed method is a Correlated Component Analysis technique. This method has low computational costs and is fully unsupervised. Moreover, we give an extension of the proposed algorithm in order to deal with a not translation invariant model. Our experimental results confirm the goodness of the method.