Texture Aided Superpixel Segmentation: Heuristic Texture Classification in the SEEDS Grid-Initialised Hill-Climbing Superpixel Generation Algorithm for Retrieval of Objects and Regions
Maciej Trzciński · Repository for Publications and Research Data (ETH Zurich) · 2013
One of the biggest challenges in Computer Vision today lies in image interpretation and particularly segmentation.Three types of cues typically used for this purpose are intensity, colour and texture.Most of the current successful segmentation techniques, based on contour detection, utilise all of them together.Superpixel segmentation algorithms however, have been so far limited to intensity and colour cues only.The difference in performance between old contour detection based techniques and the newer ones that include texture information suggests that superpixel segmentation can benefit from texture classification as well.In this work the state of the art SEEDS superpixel generation algorithm is extended with a mechanism to recognise regions with a similar spatial pattern based on the classic Laws texture descriptor.This information is then used to improve the segmentation process.Experiments were conducted on the BSDS dataset using proprietary SEEDS benchmarks and the standard precision-recall framework.Results have shown improved performance of the texture aided SEEDS approach compared to the original implementation, therefore extending the advantage over competing superpixel segmentation techniques.The promising results indicate that texture carries additional information that can be effectively used to enhance image partitioning with SEEDS.There is a measurable potential for further increase of performance while preserving the algorithm's leading time efficiency.