Color Image Segmentation Based on Learning from Spatial and Temporal Discontinuities

Abstract

This paper presents a coarse-to-fine learning method based on Extreme Learning Machine (ELM) for color image segmentation. Firstly, we locate a part of the object and background as candidate regions for sampling. By sampling from high gradient pixels (spatial discontinuity) and learning by ELM, we can extract object roughly. Due to ELM could produce different models by training with same data, the difference of their segmentation results shows some flicker in temporal feature (temporal discontinuity). So we can resampling from spatial and temporal discontinuities, and then produce a new classification model. The new model could extract object more accurately. Experimental results in natural image segmentation demonstrate the proposed scheme can reliably extract the object from the complex scenes.