Research And Implementation Of Saliency-Matting-Based Image Retrieval

With the rapid development of Internet and multimedia technology, information explosion becomes a knotty question, a big amount of information are being generated on the internet. Despite the research on text retrieval has make some big achievements, accurate image retrieval is still a big challenge. Content base image retrieval became one of the hot research fields, it is urgent to find a way to build a fast and accurate content based image retrieval system.Image matting is a technology used for splitting the foreground with the background, widely applied in image processing and video synthesis. General algorithms of image matting is supervised, additional information is needed by scribbling or color choosing to optimize the matting results. Saliency map is referred to describe the saliency of each pixel in the original image, the saliency value can be quantized into a certain range, the big the value is, the more possible the pixel belongs to the foreground. Almost all the saliency algorithms are based on one assumption that the more one pixel varies with the pixels around it, the bigger the saliency value is. Taking the saliency map as additional information, we proposed an unsupervised image matting algorithm whose performance is better than the unsupervised image matting, equivalent to the supervised image matting.In the content based image retrieval, researchers have been trying to find a kind of feature that is invariant to scale and rotation and achieved a lots. Taking sift and gist as instance, they have very nice performance in content based image retrieval. Based on sift, this paper use bag-of-features for clustering and dimension reduction, eliminates the correlation amount the features in big scale, improved the speed and accuracy. We can get the locations of the foreground objects through image matting, an overall descriptor of an image can be gained based on the location information and sift features. The descriptor can be used in image matching and image retrieval, experiments show that it has a better performance in object oriented image retrieval.Based on the algorithm proposed in this paper, a multi-feature image retrieval system was designed and implemented, besides some basic functions as images importing,image reviewing and album management, a single-feature and multi-feature image retrieval system is also implemented, the number of the images retrieved can be changed by setting the value of the similarity.