[2K1-ES-2] Machine learning: Image classification

[2K1-ES-2-02] Image selection based on autoencoder neural network and application to the semi-supervised image classification

〇Tushar Singh¹, Ashish Kumar Gaurav^1,2, Yasuhiro Tsuchida¹, Fadoua Ghourabi¹ (1. AWL inc., 2. IIT Kharagpur)

キーワード：data sampling, semi-supervised deep learning, image classification, autoencoder

Convolutional neural networks (CNNs) are becoming a key technology in processing and analyzing real-time video streams, such as security videos. When pre-processing video streams for training CNNs by splitting into image frames, we generate a large-scale image dataset from which a subset is used for training models. The random selection of a subset ignores the properties of the data and produces a repetitive dataset, which is not useful for training. This paper presents an image selection approach based on the autoencoder neural network. The autoencoder projects high-dimensional image feature vectors into a low-dimension latent space for effective analysis of image similarity. This approach allows not only to select representative images but also to facilitate the pseudo-labeling of unlabeled data. In this paper, through experiments with autoencoder, we show the benefits of this method in selecting images for training. We also explain the application to a semi-supervised image classification problem where our approach significantly enhances the accuracy comparing to random selection.