Research On Image Processing And Recognition Algorithm For Defect Appearance Of Polarizing Film

The detection method of the appearance defect of the polarizing film is still manually detected.However,the method is labor intensive,easy to be fatigued by the human eye,and reduces the detection speed of the detection of the polarizing film.In recent years,the rapid development and wide application of hardware and algorithms in the field of machine vision have laid the foundation for the automated detection of defects in the appearance of polarizing films.Our research group is devoted to the research of automatic detection technology for the defects of polarizing film appearance.From the hardware design of image acquisition to the software algorithm of image processing,all the research is gradually deepening.The paper mainly studies the image processing algorithm part.The main research contents are as follows:1.Defect detection based on improved Niblack local binarization method in spatial domain.The algorithm first accelerates the algorithm by using the integral graph algorithm,breaking through the bottleneck of the original algorithm speed.Then improve the calculation of the local binarization threshold in the mask in the original algorithm,so that the algorithm can be applied to the defect with smaller contrast,making it more robust.At the same time,by combining with the experiment,the selection criteria of the mask are designed.Finally,the value of the empirical parameter is modified so that the algorithm can be applied to the cosine background of different periods to realize the adaptation of the binarization threshold.The algorithm speeds up the operation to about 2.3 seconds,and the detection accuracy reaches about 93%.2.Defect recognition is implemented in the transform domain based on wavelet transform and frequency domain filtering.The algorithm is characterized by the fact that the defects in the image are relatively small,the background gray value is a cosine smooth transition,and the gray value of the defect is abrupt.When the image is converted to the frequency domain by DFT or DCT,The background will be a low frequency component,and the defect will be high frequency component,There is a large difference in frequency between them.Focus analysis of defect details by wavelet transform,Enhance the energy of the defect point and further increase the frequency difference between the defect point and the background.Then choose afilter with a suitable cutoff frequency,filter out the low frequency background information and only retain the foreground defect target,and finally realize the binarization through the adaptive threshold method.The running time of the algorithm is about 0.85 seconds,and the detection accuracy is about 95%,which basically meets the requirements of industrial production.3.The machine learning method based on SVM+HOG+PCA realizes defect recognition.The algorithm extracts the target region by detecting the Harris corner of the image.After Gaussian filtering and non-maximum suppression,the neighborhood range of 32*32 is selected,and the Bounding Box is finally extracted.In terms of feature extraction,Extract the HOG features of the Bounding Box,Then,the HOG features are subjected to PCA dimensionality reduction.Finally,the difficulty analysis is introduced to improve the robustness of the classifier when training the SVM classifier.Finally,use the SVM classifier classify the defect target.The running time of the algorithm is about 0.520~0.625,the recall rate reaches 0.88,the F1 value reaches 0.73,and the test set has an accuracy rate of 80%.The method is feasible.