Fabric Pilling Object Detection Based on Scale - Space Extremum

In order to solve the problem of extraction of pilling features in objective assessment of fabric pilling grading, we propose a new method for detecting pilling object using scale-space extremum. In this paper, the pilling object is modeled as an anisotropic Gaussian kernel. Based on scale-space theory and derivation of isotropic Gaussian matched filter, an operator as polynomial combinations of Gaussian derivatives is used for automatic scale selection, which provided a close approximation to Gaussian matched filter. By scale-space extrema of the normalized operator filtering, the pilling object is located and its size is measured. Depending on the anisotropic Gaussian model parameters which estimated from local structure tensor matrix, the pilling object is finally segmented and recognized. The experimental results show that the proposed method is feasible for pilling object segmentation and recognition.