Research On Key Techniques Of Eye Movement Based On Deep Learning

Eye movement technology further studies the intrinsic cognitive process of individuals by extracting data such as gaze time,number of gaze,eye distance and iris characteristics from the records of eye movement trajectories.Eye movement data classification and iris detection technology are two important branches of eye movement key technology.Eye movement data classification is widely used in market research and learning education.Iris detection is widely used in biomedical and identification fields.With the development of big data and the emergence of large-scale hardware acceleration equipment,this paper studies the eye movement data classification and iris detection algorithms respectively,and with the deep learning technology,it has achieved better results on several large-scale data sets.The main research work and innovations of this paper can be summarized as follows:(1)Many existing eye movement data classification algorithms first use artificially designed thresholds to separate eye hop data,and then use other methods to distinguish gaze and smooth trailing data.However,for low sampling rate,frame loss,general performance,eye movement data collected by eye tracker or eye movement data of different subjects,the accuracy of the classification results obtained by the threshold method is not high,especially the discrimination effect on the smooth trailing data is worse.Aiming at the shortcomings of existing eye tracking data classification algorithms,based on the in-depth study of eye movement data analysis theory and eye movement data classification algorithm,an improved self-attention based eye movement data classification algorithm(SATT-BLSTM)is proposed.The algorithm cleverly combines the self-attention mechanism with the bidirectional long-and short-term memory network(BLSTM).BLSTM captures the long motion of the eye with the input time window data feature,and the self-attention mechanism captures the intrinsic link of the BLSTM output data and further adjusts the parameters of the BLSTM.In addition,using the speed and direction of different time scales as input features further alleviates the data loss frame.The eye movement data classification algorithm proposed in this paper is evaluated on the large-scale hand-labeled dataset.The results show that the improved eye tracking data classification algorithm based on self-attention mechanism is better than other algorithms and can obtain better classification results.(2)Existing iris detection algorithms can be roughly classified into traditional and deep learning based detection algorithms.The traditional iris detection algorithm is less robust.When the resolution of the image is not high and the iris is disturbed,the recognition and positioning accuracy of the algorithm is low.The detection model based on deep learning generally has a large storage capacity and cannot Shortcomings such as real-time detection.Aiming at the defects of the existing iris detection algorithm,based on the in-depth study of the base network structure and lightweight network model of the target detection model,an improved end-to-end lightweight iris detection model MDLNet is proposed.The base network adopts a multi-granularity dense connection mechanism,which enables feature reuse and accelerates network training,and achieves better performance,smaller model storage,and faster detection speed in the case of less parameter and computational cost.MDLNet was evaluated on the Open Iris Database.The results show that MDLNet has a 58%improvement in detection speed and a 15x reduction in model size compared to the popular lightweight YOLOV3-tiny network.