Full-face Gaze Estimation Method Based On Channel And Spatial Attention Mechanism

Gaze estimation technology has a wide application prospect in human-computer interaction,human emotion analysis,commercial advertising and so on.With the development of deep learning in recent years,gaze estimation method based on deep learning has become a hot spot in the field of gaze estimation.Recent studies have shown that in the gaze estimation method based on depth learning,compared with the eye image method,the information of face image used in the full face method is helpful to improve the estimation accuracy.In order to solve the problem of large computational complexity of the existing full-face gaze estimation method based on deep learning,this thesis proposes an improved network based on an channel-spatial attention mechanism,which shows faster training speed and higher accuracy on the public dataset.The main work of this thesis is as follows:Firstly,the preprocessing method in the full-face gaze estimation is studied,and the face detection algorithm in the data preprocessing process is discussed.The true positive rate and speed are tested on the collected indoor face images.Based on the detection results of the video of the posture,the face detection algorithm based on deep learning in Open CV is comprehensively selected.Then a personal full-face gaze dataset was established.A total of 10140 full-face pictures of 3 participants were collected through the designed software and hardware platform.The head posture was free during the collection,and the gaze covered the screen.Finally,an improved network based on a channel-spatial attention mechanism is proposed.By adjusting Dense Net as the backbone network,the attention mechanism is introduced to make the network learn the importance of different channels and regions more clearly,and the channel-spatial attention mechanism is improved by changing the integration of the channel-spatial attention module from addition to element-wise multiplication,and verify the role of attention module by analyzing the distribution of attention module channels and spatial attention.Compared with the Alex Net used in the full-face gaze estimation baseline,the network has reduced the parameter amount from61.10 M to 9.75 M,and the training speed has been doubled.The leave-one-person-out evaluation is performed on the MPIIGaze dataset,compared with single eye gaze estimation method based on Le Net,the evaluation error of the network is reduced by 24.5%and compared with full-face gaze estimation method based on Alex Net is reduced by5.7%.Compared with Alex Net,the cross-dataset evaluation error on the personal gaze dataset is reduced by 11.2%.