Image Annotation Based On Multi-instance And Multi-label Learning

With the development of computer networks, storage compression technology and the popularity of smart mobile devices, visual information emerges in large numbers. How to organize, manage and retrieve massive visual information effectively becomes problem need to be solved both in scientific and commercial areas. Semantic clarity is the basis of large scale visual information management, so research on semantic annotation of visual information has important theoretical and practical significance and has attracted attention of a growing number of researchers in this area.In early time annotation on visual information is by human effort which is very time-consuming and can not meet the needs of labeling requirement in massive visual information processing, so it prompts researchers to find new tagging technology in order to achieve automatic annotation. Automatic semantic annotation of visual information is essentially a learning process from visual content to semantic tags. Because of the maturity both in theory and practical fields, machine learning techniques can provide theoretical support and possible solutions for semantic annotation, which has become mainstream programs to solve the problem of semantic annotation on visual information. Image is the commonest type of visual information and image annotation has become one of the most important and hottest direction in annotation areas.Although current visual information annotation based on machine learning has achieved some progress, there are still some problems needed to be solved, for example lacking of samples, ambiguity of visual information and mining on relevance between labels. Because multi-instance learning is the mainstream way to solve data ambiguity and image annotation is actually a kind of multi-label problem, we choose to study image annotation in multi-instance and multi-label framework. After deep study on multi-instance and multi-label framework, we introduced a new kind of label correlation to solve image annotation in multi-instance and multi-label framework. The main contents of this paper:Taking into account that traditional MIML frameworks either ignore relevance between labels or assume that label-correlation is shared by all instances in training set, we introduce local-correlation of labels and get a kind of new MIML algorithm MIML_LC_SVM. In MIML_LC_SVM we assume that label-correlation is locally meaningful which accord to the character of visual information data and experiments on Corel image set have proved the Effectiveness of this algorithm.