Face alignment is to localize multiple facial landmarks for a given facial image. While convolutional neural networks based face alignment methods have achieved superior performance in recent years, the problem remains unresolved due to the fact that L2 loss function suffers from imbalance errors of different facial components caused by region-related changes in pose, illumination and occlusions. In this situation, the L2 loss function will be dominated by errors from those facial components on which the landmarks are hard predicted. To alleviate this issue, in this paper, we propose a facial landmarks detection method based on branched convolutional neural networks, which consists of the shared layers and component-aware branches. The proposed model first captures more dependencies among facial components through the shared layers. Then, by virtue of the knowledge of component-aware branches, different sub-networks can effectively detect the facial landmarks of different components. A series of empirical experiments are carried out on a benchmark dataset across different facial variations. Compared with the existing state-of-the-arts, our proposed method not only achieves the robustness with respect to normal face and occlusion, but also effectively improves the performance of detecting landmarks on corresponding facial components.