Across-scale Local Difference L1-norm Principal Component Analysis Network for Facial Expression Recognition

Facial expression recognition has long been a research hotspot in the fields of artificial intelligence and computer vision due to its various applications. In this paper, we proposed a novel method named across-scale local difference L1-norm principal component analysis network (ALDL1-PCANet) to extract powerful and discriminative expression features. Based on the idea of PCANet model, we construct a multiscale space to calculate across-scale local differences of expression images to obtain the holist and local information. Then, we implement L1-norm PCA to learn the convolution filters of two stages from the across-scale local differences. Afterwards, we encode the output images by binary hash and concatenate all the block-wise histograms to form expression features. Finally, we employ support vector machine (SVM) with linear kernel for classification. Extensive experiments are conducted on both controlled and uncontrolled expression databases, including CK+, JAFFE, ISED and BAUM-2i. Experimental results demonstrate our proposed method outperforms the most of existing methods by effectively extracting powerful and discriminative features from both acted and spontaneous expressions.