2025 Volume 25 Issue 12 Pages 12_60-12_75
This paper presents a machine learning model that accurately discriminates whether an image frame contains an anomaly or not. It also demonstrates the feasibility of extracting the anomaly type from the anomaly content. We initially attempted to discriminate the 30 defined physical anomaly types contained in the earthquake images using the same discriminator (a CNN that discriminates 30 classes). However, we found that they could not be discriminated (correct response rate: 0.5%).We constructed an ensemble model combining six types of two-class discriminators, called weak classifiers, that are specialized for the type of abnormality to be discriminated. This allowed us to improve the final accuracy by comparing the five ensemble methods. The model with simply parallel weak classifiers showed the best discriminative performance, with a Hamming loss of 0.0015.
