Show-through cancellation in scanned documents using two-layer bidirectional neural networks

This paper addresses a problem of restoring duplex scanned documents by recovering their contents from interfering printing or handwriting on the back side caused by show-through effect. This is a nonlinear blind source separation problem. One of previous solving methods for the problem is a linear-quadratic mixing model based on a recurrent separating structure. This paper proposes a show-through cancellation model for gray scale printing. The model employs two-layer bidirectional neural networks and the networks behave stochastically and simulate the linear-quadratic mixing procedure by converting gray scale images into binary images. In addition, this paper proposes a method for suppressing blurring phenomenon by modifying the probability density function of the neuron´s output. Duplex Images with show-through processed by the two-layer bidirectional neural networks exhibit significantly lower show-through as effective as the linear-quadratic mixing model by some experiments. Furthermore, the obtained experimental results show that the method of modifying the probability density function is more effective for show-through with the blurring phenomenon.