Removal of non-stationary see-through interferences from recto-verso documents

In this paper, the removal of severe back-to-front interferences from recto and verso images of archival documents is approached from a modeling point of view, where the front and back ideal images are considered as two individual source patterns that overlap in the observed images through some parametric mixing operator. Earlier approaches were based on linear instantaneous mixing models, where the unknown model parameters were estimated along with the sources, within a blind source separation methodology. More recently, non-linear convolutional mixing models have been proposed as more realistic descriptions of the text superposition phenomenon. In particular, the model and the related restoration algorithm proposed in [1] have proved to be ecient for scans of modern documents affected by mild show-through. Nevertheless, they are not fully adequate to cope with ancient documents often degraded by the heavier and non-stationary bleed-through distortion. We then propose to modify this data model to account for non-stationarity of the degradation, and, by estimating the spatially variant model parameters from the data, we derive a computationally ecient restoration algorithm, still based on the concept of source separation. The performance of this algorithm is analyzed against documents heavily degraded by either show-through or bleed-through.