多変量ガウス分布の正規化積によるマルチビュー画像の低学習コストかつ説明可能かつスケーラブルな分類

抄録

Recently, multi-view learning, which is a deep learning method that uses multiple pieces of information about the same object to make predictions, has achieved significant success in various real-world applications. This paper proposes a multi-view image classification method that considers the uncertainty of classification probability using conventional Convolutional Neural Networks (CNN). This method is characterized by lower learning costs than other methods such as Bayesian Neural Networks (BNN) and is explainable and scalable by using the normalized product of multivariate Gaussian distributions. The proposed method can reduce the computational cost compared to the method of processing all images with CNN. Furthermore, by employing the concept of Grad-CAM, this study proposes a visualization technique for highlighting regions of interest within multi-view images. Experiments conducted with a 3-class multi-view image dataset demonstrate that the normalized product enables a rational calculation of multi-view image classification predictions and the proposed method enables a faster computation of multi-view image classification predictions than an existing method and BNN.