Image Retrieval Based On Positive Instance And Multi-instance Multi-label

With the development of science and technology, the mobile terminal equipment become more and more popular. Mobile terminals and network exchanging data all the time, which causes the explosive growth of network images. These pictures scatter in various parts of the world disorderly, but people can do limited things to analyze and dispose them. How to effectively search out what we need from these disordered images attracts people’s attention.Image retrieval technology has experienced text retrieval, content-based retrieval, multi-instance retrieval phase, etc.Multi-instance Multi-label learning has become a hotspot of machine learning research recently.In this paper we propose the improved of Multi-instance Multi-label learning and apply it to image retrieval successfully. Firstly, we introduce the development background of image retrieval,research status and related applications.Secondly, we present the algorithm and technology used in the current paper, which content Multi-instance Multi-label learning, k-nearest neighbor, K-means algorithm, MIMIBOOST algorithm, MIMLSVM algorithm and multiple kernel function.Finally, we propose a new algorithm aiming at the problems existing in the image retrieval and apply it to image retrieval, verifying the efficient of algorithm by analyzing and comparing experiment.Researchers have done a lot of work on image retrieval, however, there still exists some shortcomings. Aiming at these shortcomings, this paper make the following innovations:(1)Applying the Multi-instance Multi-label learning used in machine learning to image retrieval, consequently we propose a frame of Multi-instance Multi-label image retrieval.An image have its complete semantic information, but in the process of image tag, we use only one single category to describe the image. This will lose some important semantic information, and do bad to image retrieval. Constituting image salient area into samples by the underlying visual feature vector, thus, we can associate sample markers with images of the category. While studying the image similarity, study framework proposed in this paper is helpful to improve the precision of image retrieval.(2) In the process of image annotation, we can use the correlation between the image tags to improve the accuracy of tag, and thus excavate high-level semantic of image. There is a certain correlation between package and package, sample and sample, however, the correlation is ignored in the previous image retrieval technology. When classifying the image, it is necessary to add tags to these objects. In the process of learning between examples, some categories tag that has nothing to do with the pictures will be introduced. This can improve the precision of category labels retrieval. There is a potential semantics in the tag that between bag and bag, and the semantic relation can put a few pictures together to get its high-level semantics.(3)Proceed the underlying characteristics data of the image visual first, and then use CK_MIML algorithm to realize the classification of image retrieval. In the previous image retrieval technique, image similarity is calculated directly by using the generated package. In the fourth chapter, with the underlying visual feature vector, using the k-means clustering algorithm, we can get the clustering center. With these clustering centers, we can get the example consistency points of the positive and negative packages, then the dense point can get positive examples. Using the maximum and minimum distance of bags and clustering centers, we can constitutes the structure of the package. Using the obtained positive sample and the structure of the package, we can get a new vector to describe the images. Multiple kernel function is used to calculate the similarity between new vectors. In this way, image retrieval can be realized.