Abstract: In order to meet the speed and accuracy requirements of face key point detection (face alignment) in application scenarios, firstly, cascaded prediction is carried out on the basis of SSD (single shot multibox detector), which combines more uniformly distributed feature layers to form MR-SSD (more robust SSD), a deep learning detector with more robust response to multi-scale faces. Secondly, based on the cascade shape regression method of local binary feature (LBF), a multi-angle initialization algorithm based on the difference between the facial pixels is proposed. Five groups of feature points in the 90 degree inclination range of positive and negative face are initialized to achieve excellent fitting effect for inclined face under multi angles. The mean square deviation of each group of feature points after regression is calculated and the maximum corresponding shape is used as the final regression shape. The optimal architecture proposed in this paper can obtain robust face bounding box and face alignment schemes against multi-angle tilt in real time.