WITHDRAWN: Comparing Twins Transformer: A Steel SurfaceDefect Detection Method Based on Deep FeatureSimilarity Retrieval and Comparison

The goal of industrial quality control is to identify and detect defective components during the manufacturing process, thereby ensuring that high standards are maintained throughout the industrial production process. Existing supervised learning methods for surface defect detection rely heavily on models with optimization bottlenecks, which are similar to semantic segmentation tasks and are difficult to implement perfectly in complex defect detection tasks. Unsupervised learning algorithms are based on the modeling representation of normal samples and do not fully utilize the labeled data already obtained in industrial scenarios. In this study, we propose a feature pairing retrieval method to search for "twins" in the test samples and design a transformer combined with a Graph Convolutional Network (GCN) model for pairing discriminative defect detection, referred to as Comparing Twins Transformer (CTT). The most similar twin of the query image was used as a reference object, and the GCN was used for similarity fusion and discrimination, enabling the accurate detection of defects in the retrieved paired samples. We used multiple benchmark datasets to validate the effectiveness of the algorithm, and the experimental results showed that our method achieved state-of-the-art results for the detection of steel surface defects.