טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentOphir Boaz
SubjectCross-Talk Cancellation in Scanned Images
DepartmentDepartment of Electrical Engineering
Supervisor Professor Emeritus David Malah


Abstract

Show-Through interference is a common occurrence when scanning duplex printed documents. The back-side printing shows through the paper thus contaminating the front side image. The same occurs when scanning the reverse side of the page. This is not a problem in low quality scans (as done in home/office scanners) where, up to a degree, image quality is not an issue. The matter becomes crucial when image quality is essential. Such a case is when creating a master copy in the digital printing industry.


Previous work focused on analyzing the process that causes the phenomenon, tracking the passage of light in the scanner mechanism as it passes through the document. Obviously, any such model includes a point spread function modeling the scattering effect caused to the light as it passes through the paper blurring the back-side Show-Through. It was shown that the process can be modeled as non-linear convolutive mixture of the desired images.


In this work we present two algorithms for removing the Show-Through.


The first algorithm, improving on earlier work, attempts to alleviate the problem via an adaptive decorrelation process. A cascaded multi-stage filtering scheme minimizes the image correlation. The algorithm adapts to local brightness variations, estimating local background brightness through a mean-shift process. The effects of the cross interference, in the reference signals, are minimized by an additional post processing adaptive filtering stage.


The second algorithm treats the problem as a Blind Source Separation (BSS) problem, simultaneously estimating the images and mixing parameters. Image separation is achieved via an alternating minimization process, minimizing a cost functional combined of a Mean Squared Error fidelity term and Total-Variation (TV) regularization terms. The fidelity/regularization tradeoff is set by a location dependent scheme aiming to preserve image edges while removing unwanted show-through edges. Optimization is done via the Iterated Conditional Modes (ICM) method thus avoiding large scale optimization.


The decorrelation based algorithm achieves good results on a wide range of images at relatively low computational cost. The BSS algorithm also achieves good image separation, but at a much higher computational cost. Nevertheless, we believe that the BSS approach holds much promise for future development.