Pattern Fabric Defect Detection using Nonparametric Regression

The defect on the pattern fabric usually occurred because that pattern does not identic to the design. Therefore, in this work we proposed to solve this problem using comparing signals approach. In the first step, we modeled the images in a 2D nonparametric regression setting and let the errors of the modeled correlated in their neighborhoods. We divided the image vertically into two parts, then we construct a hypothesis that left image is the same to the right one, against they are significantly different. We then divided the image horizontally into two parts and perform the same test for the upper and lower images. To perform a test, we need to measure the distance between two nonparametric regressions and use this distance as test statistic for testing the null hypothesis. Typically, the distribution of test statistic under the hypothesis null is not known. We use standardized modification of the Mallows distance to test the hypothesis and construct spatial bootstrap to get the distribution of the test statistic. The spatial bootstrap is needed to preserve the bound of a pixel to its neighborhood