Research Of Lightweight Face Alignment Network Based On Attention Mechanism

Face alignment,or facial landmark detection is a basic challenge in computer vision,which is widely used in face reconstruction,expression classification,head posture estimation and so on.Currently,CNNs are the most effective methods for facial landmark detection.However,this works usually bring in a mass of model parameters,resulting in the training period of the model requires high computational resources.In order to achieve a balance between the detection accuracy and efficiency of the model,we introduce a multi-level attention mechanism and a high-resolution representation learning strategy to focus on lightweight face alignment algorithms in this thesis.Firstly,we introduce the Mobilnet V3 convolutional module to build the backbone network,which controls the model size to 6.5M with guaranteed detection accuracy.The new model greatly improves the detection efficiency and can be applied on portable devices with limited computational resources.In order to make the most of feature maps from different layers,the channel attention block and spatial attention block are used for the high-level feature maps and the low-level feature maps respectively.In consideration of large variation of face pose in dataset,an auxiliary network is introduced to predict the three Euler angles of face rotation and we integrate them into the loss function as weights to guide the network to aggravate the penalty for these images.The evaluation on 300 W,300VW and Menpo challenging facial landmark datasets show that the design of each module is effective.The proposed method can align some challenging face pictures accurately with so few parameters.For some images with low-resolution in the dataset,we propose a streamlined face alignment network based on high-resolution representation learning.Following the idea of the previous research point,we use HRNet V2 as baseline and introduce deep separable convolution to improve the multi-resolution convolution module in the original network.Since the shapes of real landmark heatmap pixels are different,we adopt Awing loss function to imporve the detection accuracy of model.We validate the proposed network on 300 W dataset and the results confirme the precision of proposed model.