A multi-subpopulation genetic algorithm-based CNN approach for ceramic tile defects classification

Classifying and grading the product relied on human vision has caused the poor quality products and low productivity. Due to their complicated defects, most ceramic tile factories still have relied on human vision to deal with the problem. Developing an optimal model for automatically detecting and classifying the defects is still a challenge to the companies and the researchers. Thus, this study aims to propose a multi-subpopulation genetic algorithm-based convolutional neural network (MSGA-CNN) which can automatically generate an optimal convolutional neural network (CNN) including both structure and its parameters for ceramic tile defect detection and classification based on surface images. In particular, a chromosome represents a CNN model including number of convolution layers, pooling layers, dropout rate, fully connected layers and the parameters of each layer. These structures and parameters of CNN models are optimized based on evolution processes with special encoding routine, crossover and mutation, and different selection methods. To enhance the searching ability, multi-subpopulation technique is employed in the evolution progress. In addition, a local heuristics is designed to prevent the best solution being stuck in a local optimum. A database of ceramic tile surface images was constructed for validating the proposed approach. The results have shown the efficiency of MSGA-CNN compared with other existing algorithms.