Multi-object cosegmentation using density-based clustering

Cosegmentation is one of the interesting and popular topics in computer vision. The goal of cosegmentation is to extract the common foreground objects from an image set with minimum additional information. The existing cosegmentation algorithms could be classified into two categories. One is to extract one kind of foreground objects in the image set under unsupervised approaches; the other one is to find different kinds of common foreground objects in the image set under supervised approaches of which the number of kinds should be predefined. In this paper, we propose an unsupervised cosegmentation method for multiple foreground objects, which need not preset the number of object kinds. Moreover, most of the existing cosegmentation algorithms assume that the common foreground objects should appear in all images of the image set. However, if the foreground object only appears in a few images, the object is often misclassified. Our proposed algorithm can segment different kinds of common objects and have a higher segmentation rate for some foreground objects not appearing in all images. In the proposed work, an image is considered as the combination of several objects, and each object is composed of object elements. The image set could be decomposed into lots of object elements, and then object elements with similar features could be clustered into one subobject class representing one part of an object. According to the class distribution of elements, common objects are extracted by the selection criteria. The concept of independent object elements is also proposed to increase the segmentation rate. In the experimental results, we demonstrate that the proposed approach could get better segmentation results compared with other methods. © 2020 Computer Society of the Republic of China. All rights reserved.