서강대학교

초록/요약

As the interest of the deep neural networks (DNNs) increases, applications related to convolutional neural networks (CNNs) are evolving together. This thesis proposes a novel facial expression recognition for a customized system of a limited number of people by using a cascade convolutional neural network (CCNN) with generative adversarial networks (GANs) and network compression. The process for the proposed recognition system is as follows. First, we detect the facial region on the basis of Haar-like features from individual RGB images, and the region of interest is rearranged to reduce the variation of appearance changes after GANs process and raw image process. Second, we augment the images by using GANs and various image processing techniques. Third, the augmented data trains the CCNN categorizing a person recognition and facial expression recognition (FER) in advance, and then retrains each CNN after applying network pruning to the trained CCNN. In addition, the proposed system was applied to online and offline combination method with voting and batch-size input method. In the experimental results, the proposed system had the accuracy of 99.5019 % for human classification with the loss of 0.0387 and 96.8401 % for the FER with the loss of 0.0831. Additionally, as applying network compression, human and FER classification decreased by 23.2687 % and 23.2351 % for memory resource, and by 26.5291 % and 25.8651 % for computational cost. On/offline-combined system had the accuracy of 99.0063 % for human classification with the loss of 0.0426 and 91.0476 % for FER with 0.1256. Based on this study, it could construct a personalized human-state recognition system by only using facial images in online and offline environment.