Spatial Transformer Networks For Face Alignment

Face alignment is an importent part of face recognition system.Face align-ment can effectively reduce the face pose difference through geometrical trans-formation,which can improve the robustness of face recognition system to var-ious facial pose changes.Currently face alignment method utilize facial land-mark position to align face to a predifined mean face shape.However,aligning all face to a fixed mean shape may cause the loss of face geometrical infor-mation which can help to distinguishing different faces.Especially,for some faces with large pose or rotation angle,existing face landmark detectors still have large errors and transforming these faces to front face template may bring a lot of picture distortion.In the meanwhile,it is har-d to define a single opti-mal shape for all faces.Therefore,existing face alignment methods have some limitations,specially in the face recognition scene of large pose.In this paper,we utilize the spatial transformer networks to solve the prob-lem of face alignment in the face recognition scene of large pose.We design a new spatial transformer network with cascade architecture,which can effec-tively transform face with large pose to fixed posture step by step.Meanwhile,we utilize center loss to constrain aggregated feature of different scales,which is helpful in constraining feature with rich spatial information to be more com-pact.This loss function can optimize the spatial transformer networks to abtain a more uniform alignment.We evaluate the face alignment affect and recogni-tion performance on LFW and IJB-A data benchmarks.,which proves the effec-tiveness of the method we propose.