A novel facial expression recognition is proposed in the paper, in which the local gradient dualtree complex wavelet transform dominant direction pattern is used. Firstly, four layers DTCWT are used on normalized expression image. For each layer, we can obtain the feature images of eight directions, which include 6 highfrequency directions and 2 lowfrequency directions. A new DDP (IDDP) is constructed, and which is used to code for each DTCWT feature image. Secondly, the IDDP feature images of each layer in different directions are fused based on rules of gradient direction, and every fused image is divided into several nonoverlapping and equalsized blocks. The corresponding histogram of the fused feature in each block is calculated respectively, and the final feature of facial expression image is obtained by cascading all of them. Finally, the nearest neighbor method based on Chi Square statistic weighted by Fisher is used to classify and identify. A large number of experiments show that the proposed method has a certain advantage on the recognition rate and recognition time.