Face Alignment Based On Cascade Regression Model

Face alignment aims at locating facial landmarks based on face detection results, which is the prerequisite and key for face analysis. However, face alignment is still a challenge problem due to the influences from pose, expression, occlusion and illumination. In recent years, cascade regression model has attracted widespread attention, because the shape constraint in cascade regression model is implicit, which is less affected by exaggerate expression and pose variation. As a result, the localization accuracy of cascade regression model is usually high on the wild data sets. Cascade regression model is mainly based on two aspects:local feature descriptor can describe current shape and weak cascaded regressors are able to map any complex nonlinear relationship. However, cascade regression model is sensible to initialization, the model size is too large, and it is not robust under occlusions. In order to overcome those problems, three improved models are proposed in this paper:A multi-view, multi-scale and multi-component cascade shape regression model is proposed. Firstly, we divide the training sets according to the face pose, which can decrease the shape variance within each subset, reduce the distance between the initial shape and the target shape, and accelerate the shape convergence. Secondly, multi-scale cascade shape regression is incorporated to accelerate convergence speed and avoid dropping into local optimum. Besides, multi-scale local features are utilized to implicitly incorporate local structure information, which helps to obtain better description ability. Finally, we conduct the facial component refinement according to shape variance on each facial component to provide more accurate results. Automatic face alignment results of the proposed model on 300-W 2014 challenge are exceptional.A sparse feature constrained cascade regression model is proposed. Lasso based sparse regression can select robust features and greatly compress the model size. As a result, the proposed cascade regression model is more suitable for mobile devices, which have limited computational capability and memory. Solving traditional Lasso is time-consuming, so we accelerate the sparse constraint problem by utilizing improved ALM (Augmented Lagrange Multiplier). Experimental results show that the proposed model is efficient and accurate with compressed model size.A dual sparse constrained cascade regression model is proposed. Apart from sparse feature representation, we also incorporate sparse shape constraint to improve the robustness of the model. There are two iteration steps:shape update from sparse shape indexed feature and sparse shape constraint on the update shape. Sparse feature selection is able to improve the robustness of local feature. Sparse shape constraint can suppress noises on the update shapes and accelerate shape convergence. We also accelerate sparse constraints by the improved ALM algorithm. The performance of the proposed method is better than state-of-the-art methods on the wild data sets and the proposed method significantly improved face alignment results under occlusions.In this paper, we propose three improved model to solve that the traditional cascade regression model is sensible to initialization, the model size is too large, and it is not robust under occlusions. Extensive experiments show that the proposed models achieve excellent results on many wild face alignment data sets, and surpass many state-of-the-art face alignment methods.