Research On Human Eye Vision Estimation Algorithm Based On Deep Residual Network

Human eye vision estimation is the process of predicting the direction of the line of sight.Human eye vision estimation has great research value and application value,and is widely used in neuroscience,psychology and computer science.Human eye vision estimation methods mainly include model-based visual estimation and based on apparent visual estimation.In model-based human visual estimation,complex devices(requiring multiple infrared sources and high-definition cameras)greatly limit the scope of application of this model-based approach.The appearance-based human eye vision estimation method has become a mainstream research method because of its simple equipment and wider application range.In this paper,based on the apparent line of sight estimation method,the method of learning the human visual estimation model from a large amount of data is divided into training stage and testing stage.Based the appearance-based human eye vision estimation method mainly estimates the specific position of the gaze point based on the regression method.Due to the low fitness and large errors,it is difficult to estimate the specific location of the human eye.Unlike traditional methods based on explicit human visual estimation,learning-based line-of-sight estimation requires data as a drive and can achieve good accuracy.Therefore,the learning-based visual estimation method will become the research focus of this paper.The premise of the apparent visual estimation method is to realize a high-precision human eye positioning method,and obtain the image data of the human eye through human eye positioning.The human eye localization algorithm is designed on the Kaggle dataset using the residual structure(Resnet).By observing the effect on the dataset,the performance of the Resnet-based human eye localization algorithm is very good.It also explains why Resnet is chosen as the basic network for human visual estimation.This paper first designs a human eye vision estimation algorithm based on residual structure.Since the problem often encountered in deep learning is over-fitting,this paper improves the Resnet network from two aspects.Two improvement measures are proposed.The first one is to combine the idea of the qualification trace.By improving the structure of the network and using the appropriate learning rate,the error of the improved algorithm on the MPIIGaze dataset reaches 4.06 degrees and is reduced.The second is to improve the loss function and add the corresponding trainable variables.The improved two algorithms are effectively reduced on the MPIIGaze data set.Finally,the three algorithms before and after the improvement are 60% lower than the VGGNet+ and VGGNet algorithms in the MPIIGaze dataset,and 75% lower than the KNN,RF,and ALR machine learning algorithms.The best-performing human eye localization algorithm and human visual estimation algorithm are systematically implemented on the A Natural Head Pose and Eye Gaze Dataset dataset,and finally the human visual error is below 5 degrees,which fully demonstrates the Resnet-based human eye.Visual estimation algorithm accuracy.