CNN expression recognition based on feature graph

To solve the problem of training rate decline in neural network caused by too much noise in the traditional image, a new method of expression recognition based on CNN was proposed. First, in order to narrow the face range, face image could be detected from the original image by using the AdaBoost cascade classifier. Then, the coordinates of the eye, mouth and other key parts and brow, nasolabial and other fine features could be marked by using the Harr feature and the regression tree collection algorithm. After the fusion, the feature points were generated and sent into the neural network for training. This method was tested on the 2940 face expression peak images selected from the CK + data set. Comparing to the original picture training plan, the method increased the rate by about one in a tenth.