An Efficient Facial Expression Learning Model with Video

Facial expression recognition is a challenging computer vision problem. In the process of that, the first step is face labeling, the bad effect of facial labeling will directly affect the subsequent facial expression recognition. Before, many works used MTCNN as the face preprocessing, but the accuracy has always been not good. In this paper, we propose use RetinaFace model to replace the commonly used MTCNN for face preprocessing, which greatly enhances the robustness and accuracy of the entire model; secondly, because A Compact Deep Learning Model can’t resolve multi-angle face expression recognition problem. Therefore, we propose redesign the expression recognition network by adding a dimensionality reduction module before the traditional convolution layer, which pretty reduces the thickness of the feature map and improves the accuracy of expression recognition; finally, in order to make the whole system is suitable for face recognition in video streams. Based on the problem that the number of frames sent by the camera does not match the rate of GPU processing, we propose to set the sequence cache to one, which can solve the problem of video freezing and model jumping under the condition of low-configured hardware and fluently online recognition of facial expressions. Experiments on large-scale datasets suggest that our proposed method reduces the remaining residual generalization error on the JAFFE, RER2013 and CK+datasets, reach human-level performance, and outperform state-of-the-art methods on facial expression recognition in the wild using emotion and affect concepts.