Abstract
As automation advances in manufacturing, the demand for precise andsophisticated defect detection technologies grows. Existing vision models fordefect recognition methods are insufficient for handling the complexities andvariations of defects in contemporary manufacturing settings. These modelsespecially struggle in scenarios involving limited or imbalanced defect data.In this work, we introduce MemoryMamba, a novel memory-augmented state spacemodel (SSM), designed to overcome the limitations of existing defectrecognition models. MemoryMamba integrates the state space model with thememory augmentation mechanism, enabling the system to maintain and retrieveessential defect-specific information in training. Its architecture is designedto capture dependencies and intricate defect characteristics, which are crucialfor effective defect detection. In the experiments, MemoryMamba was evaluatedacross four industrial datasets with diverse defect types and complexities. Themodel consistently outperformed other methods, demonstrating its capability toadapt to various defect recognition scenarios.