| Summary: | 碩士 === 國立東華大學 === 資訊工程學系 === 101 === Coegmentation has become a popular problem recently. The aim of cosegmentation is to segment out the similar objects from a set of images with minimum additional information. The existing cosegmentation algorithms could be classified into two categories. The first one extracts the similar object in unsupervised approach, but they could not deal with the multiple kinds of similar objects on different images. The other one proposed the multiple objects classes to find the different kinds of similar objects in the image set, but they are not unsupervised. They need to set the number of objects for segmentation. Moreover, the existing cosegmentation algorithms assume that the similar objects should appear in all images in the image set. This assumption may limit the applications of cosegmentation.
In this paper, we consider an image that is composed of several objects, like people, animal, lawn, sky etc. Each object is composed of several object elements; therefore, each image is composed of a number of object elements. Some object elements with similar features could be grouped into one object element cluster by using density-clustering algorithm. These object element clusters are the sub-object classes in the image set. Besides, density-clustering algorithm excludes a few object elements, independent object elements, which do not have sufficient number of similar object elements. Finally we de-project the sub-object classes back to images. According to the image distribution of each sub-object classes, we select some classes as the results by the selection criteria. There are three advantages of our algorithm. We propose an unsupervised multiple object class framework. We increase the segmentation rate by introducing the concept of independent object elements. We add the selection criteria for reducing the similar object constraint. If appearance count of an object class is more than the threshold, this object class is recognized as the extracted objects; even they do not appear in all images.
In the experiment, we use two types of experiments to evaluate our algorithm. The first one shows the segmentation accuracy rate compared to other approach on images set with the single kind of similar objects. Our unsupervised framework could get the same or better segmentation results. The second one is concerned about the image set which contains more than one kind of foreground objects. These different objects would not appear in all images. This experiment shows our algorithm that could reduce the similar object constraint for the image sets.
|
|---|
