Texel Discovery for Repetitive Pattern Detection

Abstract

Repetitive pattern detection (RPD) aims to localize repeated elements, known as texels. Unlike object detection tasks, RPD is challenging due to the lack of predefined classes for texels, requiring a robust algorithm that can handle arbitrary visual content and repetitive structure. One approach to solving RPD is a bottom-up method, which first identifies an initial texel and then finds the rest based on the initial one. While effective in handling complex repetitive structures, existing methods struggle to accurately determine an initial texel since they do not consider a global repetitive structure. To address this issue, we propose the self-similarity-based texel discovery algorithm (STDA) that localizes an initial texel considering its local and global repetitive structure. We observe that local and global self-similarity patterns between image patches effectively reveal a repetitive structure within a given image; local self-similarity maps reveal that all pixels within the same texel share similar local repetitive structures, while global self-similarity maps show that all pixels visually corresponding across the entire image share identical global repetitive structures. Inspired by this, our algorithm generates potential texel candidates by aggregating nearby pixels according to the local and global similarity criteria of the self-similarity map. These candidate texels are then used as templates in a template matching process to detect repetitive patterns, followed by a score-based selection process to choose the most reliable repetitive pattern. Extensive evaluations of real-world datasets demonstrate the effectiveness and robustness of our algorithm against diverse visual content and repetitive structure.