Classification of Oleaginous Yeast Images Based on Mixture of Experts with Micro-Spatial Attention

In the field of fermentation science, methods to automatically classify oleaginous yeast images are in demand. Our previously proposed method used a classification model with micro-spatial attention to focus on local spatial regions. However, this method has limitation in calculating global features, limiting the classification performance. To overcome this limitation, this paper proposes a method combining multiple classification models to capture both local and global features of oleaginous yeast images. Specifically, the proposed method constructs two types expert networks. One is to focus on local features based on our developed microspatial attention. The other is to focus on global features based on self-supervised contrastive learning. Moreover, a method for the mixture of experts integrates two classification results considering the importance of each expert network according to the input image. Experimental results for real images of Lipomyces starkeyi show that the proposed method achieved the higher classification accuracy than the existing method. Index Terms—bio-image analysis, image classification, deep learning, self-supervised learning, sparse constraint