Performance Comparison of Convolutional and Transformer Neural Networks in Defect Classification for Industrial Images

In the semiconductor industry, the utilization of advanced visual inspection technology confronts challenges due to the scarcity of the defect data and the substantial sizes of the images involved, which pose difficulties for traditional Convolutional Neural Networks (CNNs). Most CNN techniques are evaluated on common resolutions, around 224 x 224 pixels, with limited literature available on their application to larger image resolutions, such as 1024 x 1024 pixels. Furthermore, the recent surge in popularity of transformer-based models, exemplified by Vision-Transformer (ViT), underscores the evolving landscape. However, ViT's efficacy, particularly with smaller datasets (less than 2000), has prompted the development of hybrid deep learning approaches like the Compact Convolutional Transformer (CCT), specifically build to address problems related to large data. Hence, in this study, we conduct a comparative performance evaluation among ResNet50, MobileNet-v2, and CCT aiming to assess their effectiveness in handling large size images and small datasets drawn from real-world semiconductor industry scenarios. Our findings reveal that all three methods demonstrate commendable predictive capabilities, achieving accuracies of (LED Dataset [Binary Class] - Resnet50: 99.75%, MobileNet-v2: 99.87%, CCT: 97.12%; PCB Dataset [Multi Class]: Resnet50: 94.79%, MobileNet-v2: 95.67%, CCT: 94.87%) when utilizing our limited dataset sourced from actual semiconductor industry settings and publicly available resources.