Deep Learning Based Automatic Defect Detection System for Tile Walls

Quality control for tile manufacturing includes identifying defects in ceramic tiles. The tiles must be impeccable as different forms of defects-such as glaze, cratering, and cracking-weaken the quality and strength of the product. The paper introduces a novel efficient feature extraction and defect classification method on tile surfaces based on a Multi-Modal Gray- Level Co-occurrence Matrix Convolutional Neural Network. The method outdoes visual pattern and detail in capturing texture and pattern in images by efficiently combining typical image-based Convolutional Neural Networks and Gray-Level Co-occurrence Matrix-based features. The MM-GLCM-CNN system consists of two major branches-one processes information in the form of images for input in a deep CNN in which multiple convolution and pooling are utilized for extracting features and another utilizes GLCM for extracting statistical features in textures. The features are passed to a fully connected layer for complex pattern identification in information. Feature concatenation in each branch will be preceded and succeeded by dense layers and dropout for evading overfitting. The tiles are classified as belonging to one out of four-Normal, Glaze defects, Craters, and Cracks-based on output softmax classification. Experimental testing for given structure compared to classification performance in typical CNNs produced better and improved performance in classification. It includes spatial and textural features, thus providing better discrimination between subtle defects easily hard to detect based on typical techniques. The novel method introduces highly efficient automated quality inspection in tile manufacturing, reducing wastes and improving quality.