Group Image Co-Saliency Analysis Based On Multi-Instance Learning

With the rapid development of Internet technology, image data is growing at a tremendous rate. Computing resources are very limited compared with the huge amount of image data. Therefore, it is particularly important to research how to extract the more salient regions consistent with human visual perception and then give them the priority processed by computer. Image saliency detection based on human visual attention mechanism has become a hot issue in the field of computer vision.At present, the analysis on detecting salient regions among single-images has been more mature, emerging lots of effective algorithms. However, these methods are not suitable for group image co-saliency analysis. The main task of co-saliency detection is to automatically extract the common salient regions from multiple images, and the target area should be significant and reproducible. In this paper, we propose a new group image co-saliency analysis algorithm based on multi-instance learning framework. And the main research work is divided into the following two aspects:1. A multi-instance learning framework is proposed to be applied in the problem of group image co-saliency analysis. Construct the positive and negative bags with corresponding instances and then calculate the value of probability that each instance in the positive bag belonging to the co-salient regions. Based on this procedure, we transform the co-saliency analysis to a multi-instance learning problem which is simpler to solve;2. An improved sparse representation algorithm is proposed to solve the multi-instance learning problem. Construct a dictionary set and use the sparse linear combination of the dictionary set to reconstruct every instance in positive bag. In order to avoid the error that label a significant mark on the background regions with high consistency, we put forward to fuse the saliency value of every region in single-image with our algorithm to ensure a high visual saliency of co-salient regions. In addition, according to that the closer instances are, the stronger association between them is, so we regard the Euclidean distance between instances as a parameter to optimize the reconstruction process. Experiments on iCoseg and CP database show that our algorithm has a good performance.