| Summary: | 碩士 === 國立臺灣科技大學 === 材料科學與工程系 === 106 === The present fabric defect detection depends on manual examination, implemented by professional inspectors with naked eye. However, the visual detection method results in time consumption, fatigue and sub-jective defect judgment. The inspection standard is not objective enough, and tiny defects are often missed. Only about 70% of defects can be detected by inspectors. The defect detection and classification cannot be perfect. Therefore, this study designs the software and hardware equipment, combined with the developed image processing procedure to build a defect detection system for woven fabric, hoping to overcome the problems in the industry.
The detection system proposed in this study can recognize and classify eight common defects in woven fabric, including stain, broken end, broken weft, hole, nep, double pick, kinky weft and float. First, a pixels gray image is captured by CMOS industrial camera above the batcher at 20m/min. Then the defect detection and defect classification are implemented by 4-stage image processing procedure. Stage 1 is image preprocessing, the image noise is filtered by Gaussian filter. The light source is corrected to reduce the uneven brightness re-sulted from halo formation. This study uses the improved mask dif-ference algorithm to reduce the standard deviation of the corrected original image from 12.072 to 2.897. The nonuniform light source is solved effectively. Afterwards, the background texture is filtered by averaging filter, and the mean value is corrected for histogram shifting, so that this system is robust to the texture and color changes of woven fabric. The binary segmentation threshold is established by the mean value and standard deviation of image without defect. Stage 2 uses adaptive binarization to separate the background and defects and filter the noise. In Stage 3, the morphological processing is implemented before the defect contour is circled, four eigenvalues of each block, in-cluding defect area, aspect ratio of defect, average gray level of defect and the defect orientation, are calculated according to the range of contour. The image defect recognition is implemented for 2,246 images. The experimental results show that the detection success rate is 96.44%, and the false alarm rate is 3.21%. The defect classification is implemented in Stage 4. The Support Vector Machine is used for clas-sification, 230 defect images are used as training samples, and 206 are used as test samples. The experimental results show that the overall defect recognition rate is 96.60 %, providing that the software and hardware equipment designed in this study can implement defect de-tection and classification for woven fabric effectively.
|
|---|
