2018/01/14 by Ryan Webster, Webster, Ryan
Computer Science · #Advanced Image and Video Retrieval Techniques #Computer Graphics and Visualization Techniques #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Graphics (cs.GR)
paper · pdf · doi:10.48550/arxiv.1801.04619
openalex publication_date 2018/01/14 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
In this work, we present a non-parametric texture synthesis algorithm capable of producing plausible images without copying large tiles of the exemplar. We focus on a simple synthesis algorithm, where we explore two patch match heuristics; the well known Bidirectional Similarity (BS) measure and a heuristic that finds near permutations using the solution of an entropy regularized optimal transport (OT) problem. Innovative synthesis is achieved with a small patch size, where global plausibility relies on the qualities of the match. For OT, less entropic regularization also meant near permutations and more plausible images. We examine the tile maps of the synthesized images, showing that they are indeed novel superpositions of the input and contain few or no verbatim copies. Synthesis results are compared to a statistical method, namely a random convolutional network. We conclude by remarking simple algorithms using only the input image can synthesize textures decently well and call for more modest approaches in future algorithm design.