Data Reconstruction Based on Supervised Deep Auto-Encoder

Digital media information reconstruction has attracted much attention in machine learning, we propose a new method about this problem for supervising learning by using the classical unsupervised auto-encoders (AE), and we also analyze the deep model structure and training strategy. In this paper, we present a supervised deep-based auto-encoder model which has a set of progressive and interrelated learning strategies by multiple groups of supervised single-layer AE. In this structure, the one-to-one training strategy in classical AE model (one output corresponding to one input) is substituted by the multi-to-one training strategy (one output corresponding to many inputs), and it improves the ability to express the feature code. We use the structure and training strategy mentioned above to reconstruct the damaged or obscured images. Experimental results show that the proposed method has good effect and adaptability to the reconstruction of the damaged or occluded samples.