2018 年 91 巻 1 号 p. 3-8
Image classification with deep learning of convolutional neural networks (CNN) for sliced snapshots of filler morphologies in rubber materials are examined. The sliced snapshots are generated from the filler morphologies modeled by reverse Monte Carlo analysis from USAXS data of rubber composites obtained in the large synchrotron radiation facility SPring-8. In the present paper, we considered an image classification problem into 3 classes, which are fillers in the end-modified SBR and the non-modified SBR, and randomly placed points. We evaluated effects of area size and number of the used snapshots for training process on generalization ability for images which are independent from the training. The examined area sizes are (250 nm)2, (500 nm)2, and (1,000 nm)2. The examined numbers are 500, 2,000, and 8,000 per class. Here, the used architecture of CNNs consists of 4 convolution layers, 4 max-pooling layers and 3 fully-connected layers, where dropout on the fully-connected layers are used for the training. We found that the area sizes larger than (500 nm)2 and the number larger than 2,000 per class are required for good image classification. We also confirmed the generalization ability of this deep learning method is better than that of the support vector machine (SVM) for HOG (histogram of oriented gradient) descriptors.