Automatic scratch detection and inpainting

Abstract

This paper proposes a system for scratch detection using bilateral filter combined with Haar Wavelet Transform. In this system the image is first gray scaled and then bilateral filter is applied for smoothing the image as well as to retain the edges. Then the filtered image is decomposed into sub images using Haar Wavelet Transform. The decomposition of wavelet separates image into an approximated image (smoothed sub image with low frequency details) and a detail image (sub image with high-frequency details). The edge features are usually available in the high-frequency part. As scratches are generally vertical in nature the vertical detail image is alone considered to detect scratches in our work. So in this work the approximation image A and the vertical image V are taken for further processing for finding the maximum possible positions of the scratch pixels. The detection of the scratch is done in two steps. In the first step the maximum possible positions of the vertical scratches are extracted by computing means down the column using the V component. In the second step unsharp masking is used to filter A and V to enhance the edges of the image. The resultant image is converted to the binary image using an experimental threshold. Then using the computed A and V components, the image is reconstructed to get the noise free scratches. The result is further filtered using Wiener filter for removing noise completely and enhancing the resultant scratch. The resultant scratch is inpainted using the existing method in [3]. The results prove that the scratches can be detected effectively using the proposed system irrespective of illumination, color and complex background.