A method for identifying mental illness using optical flow features captured by visual sensors is proposed.The proposed method offers real-time monitoring and aids in mental illness screening at a lower cost and non-invasively.Optical flow features are extracted from facial data captured by the visual sensor and used for training TSMOSNet,which is improved by replacing normal convolution with an opti-cal flow extraction head,densely sampling optical flow feature maps,and adding temporal attention,DML distillation,and VideoMix data enhancement to improve visual features for greater accuracy.Experimental results on the H7-BDSN dataset demonstrate an accuracy of 85%and an F1 score of 0.84,which outperforms other methods.
维普
万方数据