Neurosymbolic Visual Transform Based on Logic Tensor Network for Defect Detection

Abstract

In this study, we explore the integration of visual transformers and logic tensor networks for defect detection in industrial inspections. Visual transformers excel in capturing visual patterns within high-dimensional image data, while logic tensor networks enable the incorporation of logical reasoning and domain knowledge into the learning process. This neurosymbolic AI approach enhances defect detection by leveraging visual transformer’s robust feature extraction capabilities and logic tensor networks’ ability to encode expert knowledge through logical rules. We evaluate our method on industrial MVTec AD dataset, demonstrating improved accuracy and interpretability compared to traditional deep learning models. Our findings underscore the effectiveness of combining visual transformer and logic tensor network within a neurosymbolic framework for tackling the complexities of defect detection in real-world environments. The code is available at https://github.com/YousIA/NSViT-LTN/ .