Generic colour image segmentation via multi-stage region merging

We present a non-parametric unsupervised colour image segmentation system that is fast and retains significant perceptual correspondence with the input data. The method uses a region merging approach based on statistics of growing local structures. A two-stage algorithm is employed during which neighbouring regions of homogeneity are traced using feature gradients between groups of pixels, thus giving priority to topological relations. The system finds spatially cohesive and globally salient image regions usually without losing smaller localised areas of high saliency. Unoptimised implementations of the method work nearly in real-time, handling multiple frames a second. The system is successfully applied to problems such as object detection and tracking.