Fabric Defect Detection And Classification Based On Wavelet Transformation And SVM

As an important step of quality controlling in textile companies, fabric defect detection hasbeen done manully, which causes low efficiency and suffering of workers from hard work. Therefore,this thesis designs a fabric defect detection and classification system based on wavelettransformation and SVM, which has two central modules, i.e defect detection and defectclassification.In defect detection module, this thesis mainly completes the work below:(1) Pre-processingof fabric image. This thesis uses homomorphic filtering to remove the unevenness of lightingcondition in fabric images and then enhances the contrast through histogram equalization tohighlight defect area.(2) Using multi-scale wavelet transform to find local modulus maxima todetect edge of defect area in fabric images. Considering B-spline wavelet to be smooth function,this thesis takes multi-scale wavelet transformation of fabric images, after which local modulusmaxima is calculated, then a method called modulus maxima suppression based on interpolation isused to search edge pixel of defect. Next this thesis removes pseudo edge through windowedadaptive threshold and takes image fusion of multi-scale edge images to get final defect edge image.(3) Researching defect detection based on adaptive orthotropic wavelet of fabric image. Accordingto Daubechies condition, this thesis takes the minimum approximate image energy of defect-freeimage after one-level wavelet decomposition as approximating function to get adaptive orthotropicwavelet. This thesis takes adaptive wavelet transformation of fabric images and determines the bestscale through entropy of the approximate image. After the decomposition, this thesis computes ratioof horizontal detail image and vertical detail image to be weight value, through which this thesis takes image fusion of the two detail images. Then this theis uses one-dimension maximum imageentropy to threshold the fused image to segment defect area and get the binary image which showsdefect area.In defect classification module, this thesis designs a classifier based on SVM to classify fabricdefect. Firstly, this thesis extracts features from the fused image after adaptive wavelet transform offabric image, taking the statistical properties of its gray-level co-occurrence matrix as the featurevector of training samples, training which to design1-v-1SVMs. Finally this thesis inputs the testingdefect image samples to SVM and classifies them.This thesis takes the design and implementation in software portion of fabric defect detectionand classification system. And the result indicates that this system detects the defect well and has alow ratio of wrong detection and has a relative high accuracy of classifying different kinds ofdefects.