Research On Multiple Styles Face Landmark Detection Method Based On Convolution Neural Network

With the rapid development of science and technology and the rapid development of society,facial landmark detection technology has been widely applied and gradually penetrated into people's lives.For example,face recognition,head pose estimation,face reconstruction,and 3D face reconstruction.At present,the detection of facial landmark has attracted much attention.Most of the researches mainly learn from the rich deformation of facial contours and poses,different expressions of faces,and face occlusions to obtain discriminative features.A very typical framework is to use convolutional neural networks or hand-made facial features to construct information features that describe the facial appearance and contours of the face,and then to map the information features to the positions of facial landmark through model learning.However,most face landmark detection studies ignore the problem of face image style changes.In real life,there are differences in the styles of face images we collect,light and dark,grayscale images and color images,strong and dim,and dull contrast.Changes in the face image style will also cause deviations in the detection results,reducing the accuracy of face detection.Aiming at this problem,this paper studies the methods of face landmark detection based on convolutional neural networks.The main research work and innovations of the paper include the following:First,a new type of convolutional neural network architecture is proposed.This method uses the complementary features of the original face image and the style to aggregate the face image,reduces the sensitivity of the detector to changes in the face image style,and enhances face features Robustness of point detection.Then,research and implement a landmark detection method based on multiple styles of face images.Based on the improvement of the existing CycleGAN network model,the facial feature point detection method in this paper is divided into two parts for research:1)Use the generative adversarial network to retain the input image structure and transfer the image style to generate a style-aggregated face image collection;2)Based on the style-aggregated face image collection,input the original face image again and use the style to aggregate the face The complementary advantages of the image and the original face image generate stronger predictions.Finally,experiments are designed and implemented to verify the effectiveness of the above method.The experimental results on the 300-W and AFLW datasets prove the feasibility and correctness of the method.