Face recognition based on virtual frontal view generation using LVTM with local patches clustering (パターン認識・メディア理解)

One of the major difficulties encountered by face recognition is the varying poses caused by in-depth rotations. The intra-person appearance differences caused by rotations are often larger than the inter-person differences, which makes the traditional face recognition methods such as eigen-face infeasible. This paper presents a framework for face recognition across pose based on virtual frontal view generation using Local View Transition Model(LVTM) with local patches clustering. Previous study on LVTM shows that more accurate appearance transition model can be achieved by first dividing the original face image plane into overlapping local patch regions and then the learned transition models for each patch are aggregated for the final transformation. In this paper we show that the accuracy the appearance transition model and the recognition rate can be further improved by better exploiting the inherent linear relationship between frontal-nonfrontal face image pairs. This is achieved based on the observation that variations in appearance caused by pose are closely related to the corresponding 3D face structure and intuitively frontal-nonfrontal pairs from more similar local 3D face structures should have a stronger linear relationship. For each specific location, instead of learning a common transformation as in LVTM, the corresponding local patches are first clustered based on appearance similarity distance metric and then the transition models are learned separately for each cluster. In the testing stage, each local patch for the input nonfrontal probe image is transformed using the learned local view transition model corresponding to the most visually similar cluster. The experimental results on real life face dataset demonstrate the effectiveness of the proposed method.