Face Alignment Based On Multi-scale Feature Extraction And Fusion

Due to the exaggeration of facial expressions and head poses under natural scenes or unconstrained conditions,the difference in lighting conditions and the presence of partial occlusion,the face alignment problem still faces huge challenges,so an efficient and accurate face alignment algorithm can better meet our requirements.This paper mainly from the perspective of multi-scale features,extracts more effective feature information through feature extraction and fusion while capturing the association between features,and enhances the shape constraint of the face by modeling the feature association information,thus Improve the robustness of the algorithm in complex scenes such as occlusion,lighting,and background blur.Aiming at the problem that the use of a single receptive field to extract features at each layer of the traditional hourglass network will lack the description of the overall and local associated information of landmarks,a new residual hourglass network(NRHG)is proposed,which is mainly by adding new convolution branches to increase the receptive field of the network to better extract feature information at different scales,and at the same time the size of the newly added convolution branch receptive field is adjusted for the different layers of the network to balance the relationship between the feature map resolution and the receptive field,while better retaining the structured information from the local to the whole,it also highlights the description of the local detailed feature information.Through a large number of experiments,it has also been verified that the experimental effect of the new residual hourglass network has a certain improvement over the traditional hourglass network.In response to the challenges of face alignment tasks caused by large head poses,exaggerated expressions,partial occlusion and changes in lighting in natural scenes or unconstrained scenes,we propose a mixed attention model(MAtt),which mainly focuses on the combination of modules and different levels of features in the hourglass network to capture the correlation information between features,and then modeling the association information to enhance the face shape constraints,in order to achieve a better effect of face alignment.By combining the hourglass network and the attention module in different ways,the mixed attention is further divided into two parts,global attention and local attention.The former focuses on the description of the entire face feature information,and the latter mainly enhances the expression of local features.This also makes the entire model have different degrees of attention whether it is from a local salient region or a global semantically consistent space.In addition,by adding a face pose prediction module(Hpp),combined with the corresponding category of the sample,redefining the loss function,it is helpful to deal with the problem of data imbalance,and at the same time further improve the accuracy of landmrk localization.Combined with the several key modules mentioned above,this paper implements a face alignment algorithm based on multi-scale feature extraction and fusion.Through a large number of experiments and statistics and analysis on different datasets,the effectiveness of the proposed method is proved.