부산대학교 도서관

Facial Expression Recognition (FER) plays a vital role in most of the applications in which elders and challenged persons are involved. To improve the performance of facial expression recognition (FER), this research aims to reduce the impact of redundant input from emotional-unrelated regions. In order to achieve this, a hierarchically focused densely linked convolutional neural network is suggested. In particular, Facial Landmark Detection (FLD) is used as an auxiliary task and new learning strategies are explored in a new architectural network based on Spatial Transform Network (S TN) CNN. By using it, the unwanted areas from the image are removed. Also, S TN helps for restructure the image in a proper format if the image positioning is not in proper structure. Then, the comprehensiw studies were made with the benchmark facial expression datasets FER-2013 and CK+. The result shows that the performance i.e, accuracy of the proposed system outperforms the existing methods.