The Identification Of Fabric Defects Based On Wavelet Transform And BP Neural Network

The identification of fabric defects based on digital image processing techniques can beconsidered as texture segmentation and pattern recognition problems. In this paper, theidentification of fabric defects based on discrete wavelet transform and Back-Propagationneural network is proposed, the indispensable processes of which are the defect imagespre-processing, wavelet transform, feature extraction and defects identification.During the pre-processing procedure, four different denoising methods based on wavelettransform are used to improve the peak-signal-to-noise ratio (PSNR) of the fabric defectimages. From the comparison of the 4 different denoising methods, the method basedBayesian thresholding and soft thresholding shrinkage is much better than the others.According to the special structure of the defects, a creative method is proposed to enhancethe defect images, which adopts self-adoptive Gaussian filter and Laplacian operatorconvolution and subtraction operation. With the enhancement procedure, all regular,repetitive texture patterns can be eliminated, and the defect areas become more significant.According to the special properties of the fabric defects, both textural and geometricalfeatures are extracted to improve the identification accuracy. During the textural featureextraction procedure, wavelet transform domain combined with the spatial domain methodis proposed, that is, wavelet transform combines with spatial gray level co-occurrence(SGLC) matrices methodology. The selection of optimal wavelet base, automatic selectionof multiresolution levels, reconsturution rules for subimages, automatic selection ofimportant parameters of the SGLC and the principal component analysis (PCA) of theextracted feature parameters are presented in details. The approximation condition ofselection of optimal wavelet base choosed from the wavelet base bank is the energy of thehigh frequency subimages reaches the least value. The choice of the proper number ofmultiresolution levels is reached when the energy ratio of detail subimages in twoconsecutive levels is less than 1. Subimage reconsturution rules are proposed to ensure thevalidity of the extracted features and simplify the computation at the optimaldecomposition level. The image fusion is carried out based on wavelet transform when theproper reconsturution subimages are established. The automatic selection rules fordifference distance and the direction angle of the SGLC are proposed for the continuity anddirectionality of the fabrics. The principal component analysis is executed to reduce theinformation redundancy of the 13 textural features. To synthetically present the featuresof different fabric defects, geometrical features should be extracted as well during thefeatures extraction procedure. The ratio of the maximum length and the maximum width ofthe defect is extracted from the defect image executed with the optimal thresholdsegmentation algorithm and morphology operation.The unsuitability rules of neuron number in hidden layer and training method are proposed to optimize the structure of the fabric defects identification neural network andimprove the training speed, during the identification neural network design. The 3-layerneural network is trained with the fabric defect images before identification with 7, 16and 4 neurons in the input layer, the hidden layer, the output layer respectively, and thetransfer functions for hidden layer and output layer are Logarithmic sigmoid andHyperbolic tangent sigmoid transfer function. The average recognition accuracy of defectand nondefect are 99.2% and 100% respectively, under the experimental condition, and thedefects include warp-lacking, weft-lacking, double weft, loom bar, oil stain, hole.