Overview

This dissertation, Fabric Defect Detection by Wavelet Transform and Neural Network by Tin-chi, Lee, 李天賜, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of thesis entitled Submitted by LEE Tin Chi for the degree of Master of Philosophy at The University of Hong Kong in July 2004 Textile inspection plays an important role in maintaining the quality of products. In this thesis, three methods which utilize matched masks, wavelet transform and neural network are proposed for fabric defect detection. An evaluation of the performance of the methods was conducted on eight classes of fabric defects (Broken End, Dirty Yarn, Mispick, Netting Multiples, Slack End, Thick Bar, Thin Bar, and Wrong Draw). In the first method, a multi-channel filtering bank equipped with five matched masks was used. Matched masks are 2-D filters that characterize specific texture properties. They are designed to emphasize the Wrong Draw texture, the Mispick texture, the horizontal edges, the bars structure and the filled regions on fabric images. At the filter outputs, segmentation by thresholds is applied, followed by a logical OR operation. The total number of pixels exceeding the threshold on the resulting image determines whether the fabric image is defective or defect-free. Using this method, 96% of fabric defects were successfully detected, and the false alarm rate was 6%. The method achieved a 90% - 100% detection rate for most fabric defects, though the detection rate for Thin Bar defects was only 75%. The second method employed wavelet transform to decompose fabric images into multi-scales and orientations. During the training stage, the parameters to be optimized include the rotation angles and the two thresholds applied on the horizontal and vertical transformed images. The variation in rotation angles determines the selection of wavelet bases. During the detection stage, the discrimination criterion is based on the total number of defect windows. Using this method, only 76% of fabric defects were identified, and the false alarm rate was 7%. The detection rate for Dirty Yarn was high, but much lower for Broken End and Wrong Draw defects. The last method took advantage of the fault tolerance and learning ability of neural networks. We explored the texture structure of defect-free images so that feature extraction was conducted on repeating units with proper selection of locations. For defect images, similar feature vectors were extracted and passed to the neural network. Using this method, the detection rate was as high as 92% and the false alarm rate was 6%. Dirty Yarn, Netting Multiples, Mispick, Thin Bar and Wrong Draw defects were completely identified, while 75% of Broken End and Slack End defects were detected. However, only 73% of Thin Bar defects were detected. The method employing matched masks proved the most effective in detecting fabric defects. The neural network method was next best. The wavelet transform method was the least effective, because it was only able to detect effectively certain classes of fabric defects. Dirty Yarn, Netting Multiples, Mispick and Slack End defects are relatively easy to identify correctly. Wrong Draw and Thin Bar defects are less easy to detect and Broken End and Thick Bar defects are the most difficult to detect. DOI: 10.5353/th_b2928728 Subjects: Wavelets (Mathematics)Neural networks (Computer science)Textile fabrics - Testing

Full Product Details

9781374716926

Table of Contents

Reviews

Author Information

Tab Content 6

Countries Available