papers AI Learner
The Github is limit! Click to go to the new site.

Texture-Aware Superpixel Segmentation

2019-01-30
Remi Giraud, Vinh-Thong Ta, Nicolas Papadakis, Yannick Berthoumieu

Abstract

Most superpixel methods are based on spatial and color measures at the pixel level. Therefore, they can highly fail to group pixels with similar local texture properties, and need fine parameter tuning to balance the two measures. In this paper, we address these issues with a new Texture-Aware SuperPixel (TASP) segmentation method. TASP locally adjusts its spatial regularity constraint according to the feature variance to accurately segment both smooth and textured areas. A new pixel to superpixel patch-based distance is also proposed to ensure texture homogeneity within created regions. TASP substantially outperforms the segmentation accuracy of state-of-the-art methods on both natural color and texture images.

Abstract (translated by Google)
URL

http://arxiv.org/abs/1901.11111

PDF

http://arxiv.org/pdf/1901.11111


Similar Posts

Comments