Neighbor embedding based single image super-resolution using hybrid feature and adaptive weight decay regularization
Yonggun Lee, Yoonsik Choe · 2014
In this paper, a novel single image super-resolution technique based on neighbor embedding is proposed. Conventional neighbor embedding technique uses gradient feature vector in neighbor selection. However, since gradient feature is high-frequency feature, it is not appropriate for feature representation in weak-edge and texture regions. Therefore, instead of using only gradient feature as in neighbor embedding, hybrid feature which combines gradient feature for edge region and luminance norm feature for non-edge region is used. Also, adaptive weight decay regularization is applied to shrink unstable weights for neighbors. Consequently, experimental result shows that the proposed method can reconstruct more details in weak-edges, textures, and strong-edges as well as conventional method.