Defect classification of glass substrate using deep neuro-fuzzy network with optimal parameter combination

Currently, numerous smart products are based on glass substrates. However, defects that occur during the production of glass substrates affect the quality and safety of the final products. Accordingly, we developed an optimal-parameter-combination-based deep neuro-fuzzy network (O-DNFN) for classifying defects in glass substrate images. The proposed O-DNFN comprises a deep neuro-fuzzy network (DNFN) and uses the Taguchi method. The fusion layer of the DNFN uses four feature fusion methods. The neuro-fuzzy network in the DNFN serves as replacement to a fully connected network for the classification of defects in glass substrate images. Because O-DNFN model parameter selection is challenging, we used the Taguchi method to determine the optimal parameter combination through fewer experiments. The experimental results revealed that the accuracy rates of the proposed O-DNFN with global max pooling fusion and an LeNet model in classifying defects in glass substrate images were 91.8% and 88%, respectively.