Facial Expression Recognition Based On Label Distribution Learning

Label distribution learning(LDL)is suitable for dealing with the machine learning problem with label ambiguity as same as the multi-label learning method.Label ambiguity is prone to happen in facial expression recognition due to the complexity of facial expression.So,we use LDL to solve the facial expression recognition(FER)problem better with the occasion where emotion label cannot be judge accurately.To cope the challenge of class-imbalance in FER or to improve the effect of the existing algorithm,it is essential to facilitate the learning process by exploiting correlations among labels.And high-order strategy has stronger correlation-modeling capabilities than first-order and second-order strategies.While the correlation is too complicated to be modeled,the algorithm uses high-order strategy still deliver poor performance.In this paper,two new algorithms are proposed: emotion distribution learning based on local label correlation(EDL-LLC)and label distribution learning as hierarchy combination of local models(LDL-HCLM),which both consider the high-order correlations among labels of local data-set:(1)The key idea of EDL-LLC is to balance the overall loss function and local loss function which is calculated by the sum of distance between the predicted distribution of target test sample and the ground truth distribution of every train sample.The experimental results show that EDL-LLC performs significantly better than some state-of-art multi-label algorithms and label distribution algorithms.(2)In order to cover the shortage of EDL-LLC in training time and stability,we propose a local modeling algorithm named LDL-HCLM which is based on the hierarchy combination of local label distribution models.LDL-HCLM needs to divide data-set into subsets and model sub-models like other local modeling methodologies firstly.Then LDL-HCLM will learn two types of models,each upper layer model will produces a weight distribution for its underlying sub-models.The LDL-HCLM model is constituted of the linear combination of sub-models by its upper weight distribution layer by layer.The expression ability of the LDL-HCLM model enhances as the layer of LDL-HCLM increases,so it can fit the complex data-set better.The experimental results show that LDL-HCLM has a good performance in FER.And the software based on the LDL-HCLM model achieves good performance of facial expression recognition in reality.