Fast image segmentation based on boosted random forests, integral images, and features on demand

The paper addresses the tradeoff between speed and quality of image segmentation typically found in real-time or high-throughput image analysis tasks. We propose a novel approach for high-quality image segmentation based on a rich and high-dimensional feature space and strong classifiers. To enable fast feature extraction in color images, multiple integral images are used. A decision tree based approach based on twostage Random Forest classifiers is utilized to solve several binary as well as multiclass segmentation problems. It is an intrinsic property of the tree based approach, that any decision is based on a small subset of input features only. Hence, analysis of the tree structures enables a sequential feature extraction. Runtime measurements with several real-world datasets show that the approach enables fast high-quality segmentation. Moreover, the approach can be easily used in parallel computation frameworks because calculation of integral images as well as computation of individual decisions can be done separately. Also, the number of base classifiers can be easily adapted to meet a certain time constraint.