Using teacher-student neural networks based on knowledge distillation to detect anomalous samples in the otolith images

Otoliths are small calcium carbonate structures found in the inner ear of fish and they, as one of important information carriers, are applied in diverse ecological fields. Otoliths are usually photographed and used to explore many unsolved biological and ecological questions. However, many anomalies may occur in the large volume of otolith image data due to natural or artificial consequences, which brings a huge bias to the aimed study and even misleading results. In this study, we first propose a specific definition of otolith anomalies and provide a dataset of otolith anomalies with Electrona carlsbergi, one of the most abundant species of lanternfishes, as the study subject. We modify a multiresolution knowledge distillation neural network model, the state-of-the-art anomaly detection model to a multiresolution knowledge distillation network model with asymmetric inputs, which uses grayscale maps to align the features of color maps in the feature space, to help improve otolith anomalies detection. Our fine-tuned anomaly detection network obtains a better anomaly identification performance with a Receiving Operating Characteristic Area Under the Curve value of 0.9843. Our result shown that multiresolution knowledge distillation networks can efficiently identify abnormal otolith image sample, which is of great importance for conducting otolith-based science.