An Online Multiple Instance Learning System for Semantic Image Retrieval
Chengcui Zhang, Xin Chen, Wei-Bang Chen · International Symposium on Multimedia · 2007
The proposed system is a region based semantic image retrieval system. The retrieval is realized through mapping and solving a Multiple Instance Learning (MIL) problem [1]. MIL is a supervised learning problem in which the interest is to know the instance label according to the known labels of the bag containing the instances. This problem can be mapped exactly to a region-based Content Based Image Retrieval (CBIR) scenario, in which it is assumed that the user is only interested in one particular region of the query image instead of the image as a whole. By incorporating Relevance Feedback (RF) [2] technique in CBIR, in each retrieval iteration the user is asked for the relativity of each retrieved image to the query image region. Therefore, the label of the whole image (bag) is known. The learning algorithm then needs to find out the labels of the unseen image regions (instances) in the database. In the proposed system, the One-class Support Vector Machine (SVM) [3] is applied as the learning algorithm. The overview of the system is illustrated in Figure 1.