Gaze Detection Based On Deep Neural Netword With Selective Fully Connected Layers

People's gaze contains a lot of information.Gaze can not only help machine understand people's action,ideas,and intentions,but also help to understand the surrounding environment.Gaze detection has great prospects in applications of many fields such as human-computer interaction,advertisement promotion,and visual games,but at the same time,this task is also very challenging.In recent years,deep learning technology has developed rapidly.More and more problems that seem to be difficult to solve before are solved by the penetration of deep learning technology.Applying deep learning technology to gaze detection a trend with certainty.At present,the study of introducing deep learning into gaze detection is still in its infancy and there are few research results.The gaze detection task is different from the normal single-input machine learning task.It not only needs to input the information of the entire picture,but also the information of the target person to tell the network whose gaze needs to be detected.How to properly input and process these inputs is the key to the problem.The existing gaze detection datasets are generally poor in data amount or have a poor labeling efficiency.In this paper,the data collection and labeling work was first performed,and a visual inspection data set with about 55,000 images was established.The data set has complex scenes,rich character postures.It has a certain degree of practicality.This article explores the characteristics of the gaze detection task from GazeNet,the gaze detection network proposed by the Massachusetts Institute of Technology.Experiments have shown that the line-of-sight detection network does not have to input the human face bounding box like GazeNet.Based on the summary of the existing theories,this paper proposes a method of using deep learning technology to detect the gaze of a certain person in the picture.This method only needs to input the picture information and the human eye coordinates,and extract features of the picture through ResNet.Then a special kind of network layer was introduced and named as the selective fully connected layer.It maps the position of the human eye of the target person to the part of the fully connected layer.Forward propagation is performed only in mapped area on fully connected layer.This method indirectly passes the position information of the human eye to the neural network,so that the network does not have to deal with human eye position.It reduces the number of input information,and simplifies the network structure.At the same time,for the selective fully connected layer of this paper,a new data enhancement method—data equalization—is adopted in this paper,which makes the distribution of data more uniform.Experiments show that the gaze detection network designed in this paper has achieved better results than previous studies,and had improved both in accuracy and speed.