Detection of Irregular Linear Scratches in Aged Motion Picture Frames and Restoration using Adaptive Masks

A. Machi, F. Collura, and F. Nicotra (Italy)


Image Restoration, Pattern Recognition, Digital Image Processing, Applications in Image Processing.


Linear scratches, due to abrasion occurring during shot or incorrect handling, sometimes affect motion picture films. Digital scratch restoration is normally performed by interpolating a strip region extending vertically on the entire frame and containing the scratch. In case of large scratches, some residual vertical coherence may be still perceived in the strip region even after restoration. This paper presents a method to recover detailed masks of scratched areas strictly following defect irregularities at scratch borders. The restricted mask is applied to restoration of sample actual and artificially produced film scratches. Quality restoration improvement is shown and measures of detection accuracy are reported. The detection algorithm exploits local measurements of brightness gradient. It firstly labels as scratch structure seeds picture elements where a sudden change of gradient direction occurs in the horizontal direction. Clusters of aligned seed sites are then morphologically relaxed to fully recover scratch edges and to evaluate scratch effective width. Finally edge pixels distant from their nearest scratch more than half its effective width are deleted from the mask if not topologically connected to any local short ancillary scratch structure. The algorithm is able to recover, on the average, 84% of scratch pixels manually selected, with 66% accuracy. The irregularity of mask borders positively affects the visual quality of the restored frame because it reduces the vertical residual coherence in the restored area. The local variation of the scratch width helps to recover horizontal edge structures and to reproduce local texture. The computational overhead due to the mask refinement step accounts for less than 10% of the overall detection cost.

Important Links:

Go Back