Facial expression recognition based on regional adaptive correlation

To address the problem that features extracted by CNN-based facial expression recognition (FER) do not consider structural information, a region adaptive correlation deep network (RACN) is proposed. The consists of two branches. In one branch, obtained applying CNN to sub-blocks are used as input proposed second-order (SRCN), which obtains adaptively learning regions. Furthermore, they fused with parallel branch-extracted global obtain comprehensive high-semantic feature representation. Finally, weights assigned through channel attention mechanism for more accurate classification. Experimental results show our method can effectively extract in an end-to-end manner, improve accuracy FER, and achieve competitive performance without relying on any priori knowledge. And region-adaptive extraction branch RACN be applied other networks discriminative structural-adaptive features. best knowledge, work first enrich representation static FER obtaining regional vectors via autocorrelation matrix combined compared existing literature.