Research On Multi-view Object Co-segmentation

In the information technology highly developed era, the digital image information is growing at an explosive rate. The traditional retrieval technology based on text was labor consuming and has been unable to accept, so content-based image retrieval(Content Based Image Retrieval, CBIR) technology which without human intervention has been recognized and developed. In the existing CBIR systems, most of them only process one target object image, which leads to the poor description information of the object and affects the quality of search results. In view of this situation, we put forward a novel solution, i.e. multi view object segmentation. Multi-view image segmentation can bring us two benefits:(1) segment multiple different perspective images at the same time, can rich the description information of target object;(2) according to the consistent common object in the multiple images, we can segment the object unsupervised, and release the user operation, enhance the user experience.The foundation of multi view image segmentation is single image segmentation. Owing to the multi view image segmentation requires segmenting multiple images at the same time, so we need a fast image segmentation framework in order to save time. Therefore, this article mainly focus on the following two aspects to solve the problem of multi view image segmentation:(1) Rapid image segmentation method: We proposed a fast image segmentation method based on superpixel to improve Grab Cut [1] algorithm, we use superpixels instead of pixels to simplify graph model, which can improve the running speed of the algorithm greatly. In addition, superpixel has a stronger ability to express region features than pixel can do, we select appropriate superpixel feature expression method and similarity measurement through experiment and discussion.(2) Multi-view object segmentation method: We proposed a novel multi-view object segmentation method based on mixture of links model, the method is an unsupervised segmentation method. It first obtain initial foreground samples and mixture of links model through saliency detection and image tracking, and then design a series of noise filtering and seeds propagation strategies to optimize sampling points, finally separately segment each image using our rapid single image segmentation method.In this paper, all the experiments were performed in public data sets, experimental results show that the fast single image segmentation based on superpixels not only greatly improves the speed of segmentation, but also gets a certain improvement in segmentation accuracy. The proposed multi-view image segmentation experiments show that it can release the human interaction and improve the segmentation accuracy at the same time. Compared with other multi view image segmentation method, our method has a faster processing speed while ensuring the accuracy of segmentation.