Hybrid Swin Transformer with Capsule Attention for Ceramic Tile Surface Classification

Abstract

Ceramic tile surface classification is very important for quality control since it shows how various tile types have varied textures, patterns, and surface properties. To overcome the issue, a hybrid Swin Transformer with Capsule Attention architecture is created to successfully capture both global contextual data and fine-grained spatial interactions. The architecture combines hierarchical feature extraction from the Swin Transformer with dynamic routing-based capsule attention. This makes it possible to better represent features using a dual-branch design and a feature fusion method. The VxC TSG database is used for experiments. It has several tile categories, such as Agata, Berlin, Lima, Marfil, Mediterranea, Oslo, Petra, Santiago, Somport, Vega, and Venice. Each category has three classes that show different surface conditions. The suggested model has an average accuracy of 98.83%, a precision of 98.25%, a recall of 98.25%, and an F1-score of 98.25%. This shows that it works well in all areas. Also, an AUC of 97.81% shows that the difference between things quite well. The results show that using transformer-based learning with capsule attention works well for strong and accurate classification of ceramic tile surfaces.