The Research Of Facial Landmark Detection In The Wild

Facial landmark detection is a challenging task with board applications.As the fundamental component in many analysis tasks,such as face alignment,face recongnition,pose estimation,and emotion analysis,it is also one of the most studied problems in vision.Many approaches have been proposed with varying degrees of success.Most of them work well for near-frontal faces under normal conditions but become less effective for faces that are non-frontal or under complex wild conditions like pose,lighting,expression,and occlusion.Shape regression based methods update the positions of facial landmark positions iteratively.The mean shape or shape sampled from training set is often used as the initialization,which sometimes may lead to local minimum in updation due to the offset of initial positions and target positions.Convolution neural network(CNN)based methods also introduced to slove this problem.These methods show superior accuracy,but require a complex and unwieldy architecture of deep model.To address these limitations,a joint method is peoposed in this thesis,which combine deep convolution networks with shape regression approach.Deep convolution networks in the first level predict an exact bounding box of five landmarks,the second one provide a highly robust initial shape,while the following regression finely tunes the initial prediction to achieve high accuracy.In addition,the proposed method introduces an adaptive iteration mechanism,different images have different iteration number,and output the best one among the results of all stage.Moreover,the proposed method use Lasso regression in consider the sparse characteristics.Many expreiments show that the proposed method achieves high accuracy on public datasets,especially outperforms existing methods in challenging conditions like large pose expression variation,and reduces model complexity drastically compared to state-of-the-art method based on cascaded deep model.