Spatial color histogram-based image segmentation using texture-aware region merging

We propose a new image segmentation method using spatial-color histograms that include the color and spatial information of a given image. Previous methods used a histogram with only the color information of the image or did not effectively suppress the texture components of the same object to form segmented regions, and they frequently led to the false merging of two different regions. Thus, these methods caused an over-segmentation result in the same object or an under-segmentation result in the regional boundary between two different objects. To resolve these problems, the proposed method performs a clustering that considers both color and spatial information of the image in the histogram domain and texture-aware region merging. Moreover, using a total variation-based regularizer that can remove the texture components in the same object and preserve the edge components between different objects, we improve the accuracy of region merging process that is applied to the result of the proposed histogram-based segmentation. Compared to the best results obtained using previous histogram-based methods, the proposed method achieved improvements of 0.02335 (2.910%), 0.0195 (3.977%), 0.05515 (2.431%), and 0.9639 (9.250%) in probability rand index, segmentation covering, variation of information, and boundary displacement error, which are the most widely used for segmentation evaluation metrics, respectively. Further, when compared to the state-of-the-art methods, which use the superpixel, iterative contraction and merging, and deep learning-based methods, the proposed method provides promising segmentation quality with fast operation speed.

Automated summary

The proposed image segmentation method effectively combines color and spatial information through spatial-color histograms, clustering, and texture-aware region merging, with the use of a total variation-based regularizer to improve accuracy and overcome issues of over-segmentation and boundary displacement observed in previous methods. Results show significant improvements compared to previous histogram-based methods and promising segmentation quality with fast operation speed when compared to state-of-the-art methods.