Defect detection using asymmetric Gabor filters

           One common technique to implement multi-resolution analysis is to use wavelet transforms. However, wavelet bases are shift invariant and therefore it is difficult to characterize a texture pattern from the wavelet coefficients since the wavelet descriptors depend on pattern location. Gabor filters can also decompose the image into components corresponding to different scales and orientations. Gabor filters achieve optimal joint localization in spatial and spatial frequency domain and therefore have been used extensively for texture analysis, document analysis, and object detection. Researchers have found that when a texture contains textels not arranged in square lattice , asymmetric Gabor filters are useful. The textel spacing  for real fabric samples is often not arranged in square lattice. This can be due to the presence of stochastic turf/perturbations on the fabric surface, and/or due to the underlying structure of the fabric itself as in the case of many twill weave fabrics. Thus fabric defect detection using asymmetric Gabor filters was investigated. For supervised defect detection, no effective systematic method existed previously for automatically selecting the desired Gabor filter. This work also describes a heuristic algorithm to select the best representative Gabor filter from the bank of Gabor filters to detect a class of fabric defects. The details from this work also appear in the paper from IEEE Transaction on Industry Applications.

Figure 1: Perspective view of (a) real, and (b) imaginary components of a typical Gabor filter in the spatial domain.

Figure 2: Selection of best representative filter for the supervised defect detection.

Figure 3: Plain Fabric samples with defect big-knot, wrong-draw, netting-multiplies, and mispick and their respective segmentation with the best representative Gabor filter.

Figure 4: Image histograms from two fabric defects using prior fusion scheme in (a) and (c) and the corresponding ones from the proposed fusion scheme in (b) and (d) respectively. This figure illustrates typical reduction in false alarm from the proposed fusion scheme.


 Back to research page