Estimating glue-layer defects on plywood through computer vision methods

Abstract

Making quality film-faced plywood is a vital issue for the manufacturer as it decides the quality of the end product made from this film-faced plywood. Detecting defects on the applied glue which is used to stick the films to the plywood helps in the quality control of the final products. Automation of this process decreases the time for plywood board inspection and hence helps in maintaining the quality of the product. The goal of this research was to provide a reliable and efficient technique of recognition of the glue-layer defects on the plywood. State-of-the-art convolutional neural networks and transformerbased neural networks were chosen as a method for classifying several types of glue-layer defects on plywood. In this study, four existing convolutional neural network architectures, namely VGG-19, DenseNet, EfficientNet, and ResNet-50, and two transformerbased architectures namely Swin transformer and Vision Transformer were tested for glue-layer defect identification. The Swin Transformer model achieved the highest accuracy of 88.70% followed by the Vision Transformer model, which achieved an accuracy of 88.17%. The Swin Transformer and Vision transformer models also achieved the lowest inference time of about 21 and 13 milliseconds respectively.