Diffusion-Augmented Industrial Anomaly Detection Leveraging Attention-Empowered Multi-CNN Transformer Fusion

Deep learning models have demonstrated remarkable performance in anomaly detection tasks, particularly when large datasets with sufficient anomalous samples are available. However, in industrial settings, collecting an adequate number of defective images for training remains a significant challenge. Additionally, the complexity and diversity of anomaly patterns in industrial environments often make it challenging for a single model to capture the entire range of features effectively. To address these challenges, this paper proposes the Versatile Attention Re-weighted Stable Diffusion (VARSDiffusion), a diffusion-based model for generating realistic defective samples. Furthermore, it introduces an Attention-Empowered Multi-CNN Transformer-Fused model architecture to extract diverse feature sets from anomaly images. A transformer-based meta-learner is employed to combine these features, enabling the model to detect both subtle and large-scale anomalies more effectively. The proposed approach addresses issues such as class imbalance, poor image variability, and the risk of overfitting by incorporating extensive data augmentation techniques, including the application of Gaussian noise, to improve model generalization. The model was evaluated on the MVTec LOCO AD and MVTec AD datasets, which present real-world challenges such as sparse data and imbalanced classes. The results demonstrate the robustness and reliability of the model, achieving AUROC scores of 98.15% and 97.56%, respectively, outperforming existing state-of-the-art approaches. This novel framework provides a robust and adaptable solution for industrial anomaly detection, even in environments with limited or imbalanced data.