부산대학교 도서관

In this dissertation, a fusion scheme for CNN-based universal super-resolution (SR) is proposed which works well on both face and non-face regions, simultaneously. The problems of existing SR networks are two-fold: 1) most general-purpose SR networks have poor performance on face regions. 2) face-specific trained SR networks produce better results on face regions, whereas they produce worse on non-face regions compared with general-purpose networks. To address those problems, firstly, facial feature loss is proposed, which greatly improves the perceptual quality of the faces without changing the network architecture. Secondly, a spatially adaptive fusion scheme is proposed to take advantage of both the general-purpose and face-specific networks. As qualitative and subjective evaluations, the networks with facial feature loss successfully produce meaningful facial features, such as nose shape and eyes, while the networks without facial feature loss rarely restore them. In addition, the proposed fusion scheme successfully reconstructs both face and non-face regions without boundary artifacts. In the weight-sharing experiments, it is confirmed that separately trained networks can share about 70% of the layers with little performance degradation. Therefore, the proposed scheme is expected to be an effective universal SR model which is easy to extend for various tricky objects.