Detecting Pills in Fabric Images Based on Multi-scale Matched Filtering

Abstract

The recognition and segmentation of pills in fabric images based on multi-scale matched filtering are introduced in this paper. According to the asymptotic intensity change of a pill, the pill was modeled using a Gaussian function. A matched filter is designed for detecting fabric pills. However, big pills and small pills may distribute randomly in an image and are hard to be detected simultaneously with the filter. To solve this problem, a group of matched filters, with variable-size convolution kernels using the multi-scale expansion and shrinkage properties of the wavelet function, are generated and convolved with the fabric image. Then pills, at each output of matched filtering, are recognized and we determined whether pills exist in the local area by merging all of the outputs of the multi-scale matched filtering. Finally, pills in the local area, where the existence of pills has been confirmed, are further segmented using an adaptive threshold method based on mean and weighted variance. The detection of pills on a real image shows that this method can satisfy the recognition and segmentation of different size pills and is helpful for quantitatively describing the pilling information.