The Detection And Identification Of Fabric Defects

With the continuous improvement of people's standard of living,the requirement on the quality of fabrics has become increasingly strict.According to the survey,the reasons for the poor quality of clothing in the apparel industry are mainly related to fabric defects,so detecting and discriminating fabric defects is a crucial link in fabric quality inspection.After the fabric defect image is subjected to discrete Fourier transform,the corresponding spectrum diagram will be obtained,the bright spots on the spectrogram contain fabric defect information.Based on this,this thesis proposes a fabric defect detection algorithm in the frequency domain.BP neural network has extremely strong nonlinear fitting ability which is often used in pattern recognition.In this thesis,BP neural network and random forest are combined,and the combined network is used to identify defects in fabric images.(1)For the spectrogram of the fabric defect image,this thesis designs a frequency domain filter,which contains multiple optimal circular frequency domain Gabor filters for covering the bright spots around the center bright spot of the spectrogram,and one for covering the spectral chart the center mask of the frequency domain of the center bright spot.According to the difference in the number of optimal Gabor filters selected,this thesis proposes two frequency domain fabric defect detection algorithms—algorithm one and algorithm two,and uses two algorithms to detect fabric defect images.Experiment shows that two algorithms can effectively detect different types of defects,but the detection effect of the algorithm one on a pair of defect details is unsatisfactory,while the algorithm two has a better detection effect on the defect details.Finally,the algorithm two in this thesis is compared with other algorithms in terms of algorithm and detection effect.The result shows that the defects detected by algorithm two are closer to the real defects.(2)In order to judge whether the fabric image contains defects,this thesis designs a network based on BP neural network and random forest for discrimination.The network extracts the feature vectors of the three fabric images with the help of spectrogram filtering,gray level co-occurrence matrix and local binary pattern(LBP)operator,then,training and testing by three independent BP neural networks.Last,the random forest classifier synthesizes the test results of three BP neural networks to get the final discriminant result.Experimental result shows that the network designed in this thesis has strong discriminating ability and the discriminating accuracy rate reaches 99.9%.Meanwhile,we also analyzed the reasons for the misdetection of three fabric images with high frequency of misdetection.Finally,comparing the network of this thesis with other networks from the perspective of feature vector extraction method and network complexity,the results show that this network is simpler and more efficient.