MTFM: an unsupervised multiteacher feature matching method for photovoltaic cell anomaly detection

Photovoltaic cells are essential components for solar power generation. Anomaly detection in photovoltaic cell images is a challenging task due to the difficulty in obtaining defect samples in industrial production, variations in abnormal region sizes. Prior methods utilizing single-teacher networks suffer from under-learning in this scenario. To address these issues, this paper proposes a Multi-Teacher Feature Matching (MTFM) methods for anomaly detection. In this approach, we employ multiple powerful teacher pre-trained networks to guide the student network in learning representations of normal samples. By assigning different weights to the teacher networks, the student network focuses on learning representations from teachers with strong learning capabilities. Furthermore, a multi-teacher multi-scale abnormality map fusion strategy is introduced, enabling the comprehensive integration of different scale abnormality information from various teacher networks, thus improving the detection of anomalies at different scales. We utilize the discrepancy of feature distributions between the multiple teacher networks and the student network as an anomaly scoring function to represent the likelihood of anomaly occurrences. Experimental results on the MVTEC and PVEL_AD datasets demonstrate that our proposed method significantly enhances knowledge transfer between teacher networks and the student network, leading to efficient image-level anomaly detection.