An Expression Recognition Transfer Method Of Extended Facial Landmarks Based On Facial Feature Space

In recent years,facial landmark detection technology has been widely used in facial expression recognition methods and has made outstanding progress.However,the targets and datasets of landmark detection are different from expression recognition.In order to inherit facial landmark information effectively,it is necessary to use a transfer learning method to transfer facial landmark features to the expression recognition network.In addition,improving facial landmark detection technology can improve the effect of facial expression recognition methods.To this end,this paper proposes a transfer learning network(EFLTNet)from a facial landmark detection framework to an expression recognition framework and designs an extended facial landmark dataset(EFL)for expression recognition tasks based on facial landmark detection.In the heatmap regression method of the core module of facial landmark detection,this paper proposes the information transmission retention model(ITR)for the first time to ensure that facial feature spatial information is retained and integrated during the heatmap generation process;in addition,this paper innovatively proposes a heatmap classification network of heatmap regression method,which can reduce the error of heatmap mapping.Experiments have proved that for expression recognition tasks without pre-training,the transfer effect of using 50 EFL datasets exceeds that of the WFLW full dataset.This paper verifies the effect of EFLTNet on the RAF-DB expression recognition dataset,and finds that it can achieve comparable results compared with the state-of-the-art SCN method on the full dataset;as the size of the training dataset declines,it achieves 62.78% accuracy on 4000 images and 53.00%accuracy on 2000 images,which are 12.16% and 14.31% higher than by using the stateof-the-art SCN method respectively.