A Back-Propagation Neural Network for Recognizing Fabric Defects

Abstract

Appearance is an important property of fabrics. Traditionally, fabric ins on is done by workers, but it is so subjective that accuracy is a problem because inspectors tire easily and suffer eyestrain. To overcome these disadvantages, an image system is used as the detecting tool in this paper. A plain white fabric is adopted as the sample, and the distinguishing defects are holes, oil stains, warp-lacking, and weft-lacking. An area scan camera with 512 × 512 resolution is used in the scheme, and a grabbed image is transmitted to a computer for filtering and thresholding. The corresponding image data are then used in the back-propagation neural network as input. There are three input units, maximum length, maximum width, and gray level of fabric defects, in the input layer of the neural network. This system is successfully employed to determine nonlinear properties and enhance recognition.