Efficient Cosegmentation In Internet Images Collections

Image segmentation has long been a fundamental and challenging task in computer vision. Traditional methods are focused on single image segmentation by utilizing some clues such as color difference and sharp edges to get the local coherent area. However, segmenting by one single image is a challenging problem and sometimes we can’t get satisfactory results without prior information and user interaction ([1][2][3][4]). Then co-segmentation was proposed to jointly segment multiple images containing "similarly looking objects", which can serve as a means of compensating lack of priori. Co-segmentation could segment objects with less or no user workload compared with traditional methods. Hence co-segmentation can be applied to many applications, such as mining vision information from photosharing websites like Facebook(consumers always upload several related pictures like self-portrait)[5], and easily getting the common objects of Internet image collections returned by image search engine for a user query[6].However, when it comes to segment Internet images jointly, we find there may be a number of noise images (not contain objects of interest or objects are not the main content) mixed in the dataset. Most existing methods of co-segmentation are based on the assumption that every image contains the object of interest, thus they can’t handle noise images effectively. Then In the work of [6], Michael Rubinstein first presented a method which could discover the common objects of Internet images. His method performs well in the dataset containing different style, color and viewpoint common objects. However, this method is time-consuming and could also fail in some cases. In this paper, our goal is to handle noise images and segment out final common objects efficiently. Considering computational cost of complex features and great developing of image search technology, which means searching result is more uniform, we shift our focus to the simple images features to achieve co-segmentation efficiently. The experimental results demonstrate our method performs well in many image groups with lower computational cost.