Lite AU convolution network driven by a small amount of samples

Abstract

A novel lite action unit (AU) convolution network (LAUCN) is proposed for automatic AU detection, which could improve the accuracy of AU detection with a few samples. (i) LAUCN could transform the manual intervened factors (i.e. gradient, pixels etc.), which identify the characteristics of two-dimensional shapes, to a hierarchical perceptron in multi-direction and multi-scale, while a small amount of samples are needed. (ii) LAUCN takes into account the core properties of facial AUs for detecting action unit more accurately, where the existence of strong co-occurrence structure indicates that the presence of one AU can be regarded as a prior for the presence of others so that the problem of subtle changes in appearance or geometry of the face will be well solved via exploiting this property. The experiments are conducted on the CK+ dataset, and the experimental results show that LAUCN can accomplish AU recognition satisfactorily and has achieved the expected results as well.