Abstract:  Traditional human abnormal emotion recognition algorithms have a high error rate in recognition due to the influence of noise. To solve this problem， a human abnormal emotion recognition algorithm based on video images is studied. Firstly， the facial expressions and body movement features were detected to obtain feature vectors through image preprocessing， and the target image classification features were obtained by scanning pixels. Deep learning was used to extract facial features， and a MLP hierarchical architecture was applied with corresponding learning rates and iteration times to obtain a two-dimensional training dataset of images.Then， the feature maps were sent to the network connection layer for dimensionality reduction， and the classification layer of the model was set. Finally， feature transfer and effective recognition were achieved by connecting the feature maps of different layers through DNET. The experimental results showed that the recognition error rate of the proposed method was low，and the recognition effect of the test set after adding noise processing was significant， which could efficiently and accurately recognize human abnormal emotions. 

Key words: video images, human abnormal emotion, recognition, deep learning