Two-layer features extracted from essential areas for facial expression recognition

This work presents a novel two-layer feature extraction method on essential facial areas. First, each original image is divided into three local images (first-layer area) according to the facial organs carried significant information. Second, the local images are down-sampled to images of lower resolutions (second-layer area). Then the texture and frequency features are individually extracted by the Local Binary Pattern (LBP) and Discrete Cosine Transform (DCT) operators. Finally, the Support Vector Machine (SVM) is adopted to classify seven facial expressions with Japanese female facial expression (JAFFE) database. Experimental results shows that the second layer of the proposed scheme can significantly boost the recognition rate by around 10%. As a result, the proposed scheme can obtain the highest recognition rate among various former methods. Moreover, the proposed method can achieve a better recognition rates with a lower resolution than those methods in the literature with higher resolutions.