Deep auto-encoder based on supervised learning for damaged face reconstruction

Abstract

Based on the reconstruction idea of auto-encoder (AE) and image reconstruction, we present a new idea that the classical auto-encoder can be polished up by supervised learning. We also present a novel supervised deep learning framework for damaged face reconstruction after analyzing the deep model structure. The proposed model is unlike the classical auto-encoder which is unsupervised learning. In this paper, the deep supervised auto-encoder model is illustrated, which has a set of “progressive” and “interrelated” learning strategies by multiple groups of supervised single-layer AE. In this structure, we define a Deep Supervised Network with the supervised auto-encoder which is trained to extract characteristic features from damaged images and reconstruct the corresponding similar facial images, and it improves the ability to express the feature code. Extensive experiment on AR database demonstrates that the proposed method can significantly improve the smoothness of the damaged face reconstruction under enormous illumination, expression change. Experiments show that the proposed method has good contribution and adaptability to the damaged face reconstruction.