文書画像の解析に適した文字認識AutoEncoder の検討

抄録

AutoEncoder, which acquires a specific feature space model from unsupervised data, has come to be one of the key technologies for designing a system based on neural networks. In this paper, we conduct an assessment of three types of constraint for AutoEncoder. As results of two experiments, we confirmed the sparse coding constraints is valuable for applying the acquired feature space to noisy photographed document analysis.