Abstract: For the problem that autoencoder can not obtain class information according to labels during the unsupervised training process, to improve the recognition accuracy, stacked class denoising autoencoder(SCDAE) is proposed to extract class information, and autoencoder combination features extraction method is adapted to extract combination features for classification. The method builds a stacked denoising autoencoder(SDAE) and a SCDAE. The method fine-tunes SDAE and SCDAE to form a combined model (CM). The combination features containing main information of the input data and class information are acquired through CM, and they are further used for classification. MNIST and USPS handwritten database are selected for testing the proposed method, and the results show its superiority on extracting features and improving recognition accuracy.

Key words: stacked denoising autoencoder, stacked class denoising autoencoder, class information, combination features