DeepProp: Extracting Deep Features from a Single Image for Edit Propagation

• DOI: 10.1111/cgf.12822
• Publication Date: 2016-05-27T15:11:01Z
• Authors: Yuki Endo; Satoshi Iizuka; Yoshihiro Kanamori; Jun Mitani
• Keywords: Feature (linguistics); Similarity (geometry); Representation; Feature Learning

Abstract Edit propagation is a technique that can propagate various image edits (e.g., colorization and recoloring) performed via user strokes to the entire based on similarity of features. In most previous work, users must manually determine importance each feature color, coordinates, textures) in accordance with their needs target images. We focus representation learning automatically learns representations only from single instead tuning existing features manually. To this end, paper proposes an edit method using deep neural network (DNN). Our DNN, which consists several layers such as convolutional combiner, extracts stroke‐adapted visual spatial features, then adjusts them. also develop algorithm for our DNN does not suffer vanishing gradient problem, hence avoids falling into undesirable locally optimal solutions. demonstrate without manual tuning, achieve better results than work.