(英) |
This paper presents an anomaly detection method for estimation of deteriorated regions from rubber material electron microscope images. In order to develop rubber materials with high durability, it is important to clarify the cause of deterioration. For analyzing the cause of deterioration, it is expected to utilize machine learning technology, especially deep learning. Although deterioration of the rubber materials can be observed from electron microscope images, it is difficult to obtain a large number of deteriorated data. Hence, we solve the above problem by using feature representations based on deep learning. Deep convolutional neural network (DCNN) can learn high representation features from target data sources, and extracted features from pre-trained DCNNs have been used by many researchers. In this paper, we can obtain features based on deep learning from rubber materials by using such pre-trained DCNNs. Finally, we can estimate deteriorated regions based on anomaly detection by using the obtained features. In this paper, we verify feature representations extracted from DCNN models to improve the estimation performance. |