Presentation Information
[2E5-GS-10o-04]An Ensemble-based Classification Method for Image Recognition with Limited Stag Beetle Data
〇Kazuma Murakami Murakami1, Naoki Mori1, Yusuke Kuroda2, Hiroki Gotoh2 (1. Osaka Metropolitan University, 2. Shizuoka University)
Keywords:
Bioimage Informatics,Fine-grained Image Classification,Limited Data,Ensemble Learning
Recent progress in image recognition has made fine-grained image classification more practical. This task aims to distinguish closely related taxa based on small morphological differences. In this study, we focus on Serrognathus titanus, a stag beetle species for which genetic disturbance is a concern. Our goal is to identify regional populations within the same subspecies using images of mandibles.
At this taxonomic level, the number of available samples is usually limited. When the dataset is simply split, rare morphological features may not be included in the training data. To address this problem, the dataset was divided into six subsets. One subset was fixed as test data, and the remaining five subsets were used alternately as validation data. In this way, all data except the test data were used for training. As a result, five different training datasets and five corresponding models were created.
For each model, inference results on the test data were analyzed using Attention Rollout to visualize the basis of the model’s decisions. In addition, we introduced a dummy class to enhance the diversity of each model. This led to higher accuracy compared to the setting without the dummy class.
At this taxonomic level, the number of available samples is usually limited. When the dataset is simply split, rare morphological features may not be included in the training data. To address this problem, the dataset was divided into six subsets. One subset was fixed as test data, and the remaining five subsets were used alternately as validation data. In this way, all data except the test data were used for training. As a result, five different training datasets and five corresponding models were created.
For each model, inference results on the test data were analyzed using Attention Rollout to visualize the basis of the model’s decisions. In addition, we introduced a dummy class to enhance the diversity of each model. This led to higher accuracy compared to the setting without the dummy class.
Comment
To browse or post comments, you must log in.Log in