Multi-model anomaly detection for industrial inspection with dynamic loss weighting and soft-hard features loss

Anomaly detection in manufacturing remains a significant challenge, particularly due to the unique characteristics of industrial data and the scarcity of abnormal samples. This study investigates the effectiveness of unsupervised learning approaches, leveraging multi-model frameworks that combine pre-trained models on large-scale datasets with models specifically trained to capture industrial data features. This study proposes two key methodologies to maximize multi-model efficiency. Dynamic loss weighting optimizes the contribution of each model during training, enabling networks with diverse expertise to synergize effectively. Soft-hard feature loss, in particular, focuses on precisely capturing subtle anomaly regions that traditional methods might miss. By emphasizing features with high-error values while appropriately utilizing those with lower error values, the proposed approach enables more detailed anomaly detection compared to existing methods, allowing for the detection of even minor defects through refined anomaly region analysis. Quantitative results on the MVTec and VisA datasets demonstrate that the proposed method achieves remarkable performance improvements, with up to +0.6% (AU-PRO) on the MVTec AD dataset and up to +0.4% (AU-ROC) on the VisA dataset compared to baseline methods. In addition to its superior quantitative performance, the proposed method enables more precise anomaly detection than conventional approaches. Furthermore, the proposed method eliminates the need for experimental tuning, enabling its application to diverse multi-model approaches and ensuring adaptability across various datasets.