Image quilting for texture synthesis and transfer.pdfVIP

Image Quilting for Texture Synthesis and Transfer Alexei A. Efros William T. Freeman Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, Massachusetts 02139 @ MIT The Problem: A texture synthesis algorithm should be able to take a sample of texture and generate an unlimited amount of image data which, while not exactly like the original, will be perceived by humans to be the same texture. Furthermore, it would be useful to be able transfer the texture from one object to anther (e.g. the ability to cut and paste material properties on arbitrary objects). In this work we present an extremely simple algorithm for texture synthesis. We synthesize new texture by taking patches of existing texture and stitching them together in a consistent way. Figure 1: Demonstration of quilting for texture synthesis and texture transfer. Using the rice texture (upper left), we can synthesize more such texture (upper right). We can also transfer the rice texture onto another image (lower left) for a strikingly different result. Previous Work: With Heeger and Bergen [4] the quality of texture synthesis reached a level acceptable for use in computer graphics. Their main insight was to model texture in terms of histograms of filter responses at multiple scales and orientations. But these marginal statistics fail to capture important relationships across scales and ori- entations for structured textures. Several attempts have been made to extend this idea to capture a wider range of textures, including De Bonet [1] who samples from conditional distribution over multiple scales, and later Por- tilla et.al. [5] who match both first and second order properties of wavelet coefficients. While important from a theoretical point of view, neither method is successful at capturing local detail of