A new semi-supervised learning for image classification

Expansion of applicable scope of deep learning

Abstract

Since image annotation work in image classification is expensive, the number of labeled images is small compared to the number of unlabeled images. In this paper, we propose a new unsupervised learning ASL (auto-similarity learning) in order to effectively use large amounts of unlabeled images. ASL learns image similarity between real image and masked image by generating multiple masked images in which some pixels are replaced with random numbers from pixel values of real image. By simultaneously executing ASL and ‘supervised learning using labeled image', we construct semi - supervised learning which realizes a neural network with high generalization.