Abstract: Image segmentation is an important issue on image analysis and understanding,which means that the given image is segmented into some non-overlapping regions and each region has the same or similar intrinsic properties.These intrinsic properties are agreed in the same region,whilst they are different between different regions.In many ima-ge processing applications,due to its large amounts of image pixels,some pixel-based algorithms are always time-consuming and memory-demanding.The image superpixel-based scheme will be an efficient solution to alleviate the storage and time complexities.Based on the analysis of regional covariance,this paper presented a novel similarity measure for image regions and a robust scheme for generating image superpixels.Firstly,the input image is divided into some small regions using K-means algorithm,and then the intrinsic image properties are described by high-dimensional regional covariance matrix.Then,the similarity measure of different regions is determined by the regional covariance distance.Finally,combining the Graph-based scheme and K-means clustering,the final image superpixels are generated.Compared with other superpixel generation approaches,our proposed method is efficient and can reduce some unnecessary over-segmentation.Our algorithm can also keep the image edge information and reduce under-segmentation errors for generating compact image superpixels. The superpixel generation scheme can be applied to the stylized rendering,which will lead to the artistic oil painting.

Key words: Image segmentation,Superpixels,Region covariance,Clustering,Stylized rendering