Research On Facial Landmark Detection Algorithm Based On AC-SE-ResNeXt Convolution Neural Network

In recent years,with the rapid improvement of computer computing power and the continuous deepening of deep learning algorithm research,facial landmark detection algorithm has also been developed by leaps and bounds.Due to its important application in facial expression recognition,face tracking,identity verification and other scenes,it has been the focus of attention and research hotspot in the field of computer vision.Among the existing classical facial landmark detection algorithms,convolutional neural network based on cascade regression is an important branch.However,the strategy has the following shortcomings with high accuracy:(1)due to the idea of cascade regression,the accurate key point position is obtained step by step from coarse to fine,so multi-stage regression is needed.(2)In order to reduce the input size of the next stage and reduce the noise impact of the input data,it is necessary to preprocess the input data between adjacent stages.(3)The algorithm flow structure is quite complex,which leads to the complexity of these algorithms in both the initial training stage and the application stage.To solve the above problems,an end-to-end new network structure is constructed with attention mechanism based on the residual network to achieve the accuracy not less than cascade algorithm.The main innovations of this paper is as follows:1.Based on the Aggregated Residual Transformations for Deep Neural Networks(ResNeXt),the attention mechanism is introduced by reconstruct ResNeXt model.Where the spatial attention module is constructed with Asymmetric Convolution(AC)structure,and the channel attention module is generated with Squeeze Excitation(SE)structure.The serial network and parallel network of Asymmetric Convolution-Squeeze Excitation ResNeXt(AC-SE-ResNeXt)model are designed and constructed by strengthening the learning ability of network space and the channel correlation of feature mapping.2.Above AC-SE-ResNeXt model with AC structure and SE structure can realize the end-to-end processing.Moreover,using the unitary network with single stage,the complexity of multi-stage regression algorithm in cascade strategy is reduced greatly,and the procedures of data preprocessing between adjacent stages are obsoleted.3.The trained models are tested on BioID and LFPW datasets respectively.The average error rate of five key points detection is 1.99% on BioID datasets and 2.3% on LFPW datasets.The experimental results show that,compared with the cascade algorithm,it can not only simplify the flow structure of the algorithm,but provide the end-to-end processing as well.Meanwhile,the accuracy is not less than the old one and the robustness has been improved significantly.