Robust defect detection in plain and twill fabric using directional Bollinger bands

Abstract

A directional Bollinger bands (BB) method for the detection of defects in plain and twill fabric is presented, whereas a previous BB method was for patterned Jacquard fabric. BB are constructed using the moving average and standard deviation to characterize any irregularities (i.e., defects) in a patterned texture. Every patterned texture constitutes a primitive unit that can be used to generate the texture by a translational rule. The regularity property for a patterned texture can be implicitly regarded as the periodic signals on the rows and columns of an image. To utilize such a regularity property, an embedded shift-invariant characteristic of BB is explored. The original BB method is further developed using directional rotation iterations, which enables the detection of directional defects in plain and twill fabric. The directional BB method is an efficient, fast, and shift-invariant approach that enables defective regions to be clearly outlined. This approach is also immune to the alignment problem that often arises in the original method. The detection accuracies for 77 defective images and 100 defect-free images are 96.1% and 96%, respectively. In a pixel-to-pixel evaluation comparing the detection results of the defective images with the ground-truth images, a 93.51% detection success rate is achieved.