Defect detection of LGP based on combined classifier with dynamic weights

Abstract A novel method based on machine vision for defect detection of light guide plate (LPG) was proposed in this study. After image segmentation and enhancement, the proposed combined classifier with dynamic weights (CCDW) was trained to classify the LPG samples considering both feature extraction diversity and base classifiers diversity. First of all, two categories feature extraction was used to improve the feature extraction diversity. The former (manual extraction) included geometric characteristics, Hu moment, and histogram of oriented gradient (HOG); the latter (automatic extraction) was executed by deep neural network (DNN), such as convolution neural network with two convolution layers (CNN-2C) or with three convolutional layers (CNN-3C), recurrent neural network (RNN) and residual neural network (ResNet). Moreover, the proposed CCDW, considering base classifiers diversity, selected the best combination of features for discriminating between a defective LGP sample and a good one. Experiment proved that our proposed CCDW method outperformed mainstream methods, such as AdaBoost classifier, decision tree classifier (DTC), CNN, ResNet, etc. It achieved the desired prediction rate of 0.9161 and f-measure of 0.8906 with small standard deviation (0.0128 and 0.0141, respectively). Furthermore, the strategy of applying dynamic weights was better than that applies fixed weights in the proposed combined classifier. Hence, selection of dynamic weight rules for combined classifiers was below: dynamic weight of three better base classifiers (AdaBoost, CNN-3C, and ResNet) will increase from 1/8 to 1/4 when their predictive probability is good; dynamic weight of five worse base classifiers (DTC, K-Nearest Neighbor (KNN), random forest classifier (RFC), CNN-2C, and RNN will decrease from 1/8 to 1/16 when their predictive probability is bad. Lastly, the high threshold (0.65) of proposed CCDW can boost the maximum capability of making judgment.