Multi-class Object Recognition and Segmentation Based on Multi-feature Fusion Modeling

Abstract

The paper presents a new theme and region based CRF model to realize the combination of multiple image texture, shape, context and location features. The model parameters are learned by Joint-boosting algorithm. The over-segmentation algorithm is used to divide the image into finite continuous regions. The constraint relationship between image theme, region and pixel is considered while modeling feature potentials and optimizing parameter's selection to improve the accuracy of multi-class object recognition and segmentation. The experimental results on MRSC-21 database show that the accuracy of the algorithms proposed in this paper outperforms that of the other existing algorithms. Especially by concerning regions and theme factors, our model obtains improved accuracy of segmentation and recognition of highly structured classes of objects with large shape variance and fewer training examples.