Research On Joint Dictionary Learning Model Based Feature Space For Canonical Domain Expression Recognition

The technology of facial expression recognition has become an important topic due to its numerous applications.Although the existing facial expression recognition algorithms have achieved significant recognition performance,high-precision recognition of small-scale expression samples is still challenging in facial expression recognition task.Therefore,the paper researches from the perspective of feature extraction and constructs joint dictionary learning model based feature space for canonical domain expression recognition according to the characteristic of small-scale expression samples.The detailed research contents are as follows:Firstly,according to insufficient intra-class expression variation and weakened inter-expression correlation caused by small samples and the correlation between different expression is always ignored,multi-class differentiation feature representation guided joint dictionary learning for expression recognition is researched.The multi-class differentiation feature dictionaries are constructed by multi-class sparse reconstruction,aiming to increase the linear separability among expression samples.Then the collaborative matrix of multiple feature dictionaries is jointly learned,aimed at establishing the strong representation relationship among multiple feature dictionaries.Secondly,given the outstanding feature representation ability of the deep subspace model and diverse representation of dictionary atoms,fusing the feature representation based on incremental kernel principal component analysis network(IKPCANet)and jointly constraint dictionary learning is researched.The multi-class filtered feature dictionaries are constructed by means of IKPCANet,with the aim of enhancing the discriminative representation capability of expression features.In addition,the compact constraint between atoms in the multi-class filtered feature dictionary is jointly constrained by introducing the locally compact constraint distance,aimed at improving the discriminative performance of the learned dictionary.Finally,according to single feature is difficult to more comprehensive express the limitation of small-scale expression samples,dual-stream feature fusion based prior feature and deep feature for expression recognition is researched.The algorithm integrates prior feature and deep feature to obtain more comprehensive and more discriminative expression feature,aiming to make up the influence caused by single feature.Then the algorithm uses channel-attention module to focus on discriminative features that contribute to classification in dual-stream fusion features and improves expression recognition performance.