Image classification with multi-view multi-instance metric learning

Abstract

Image classification is a critical and meaningful task in image retrieval, recognition and object detection. In this paper, three-side efforts are taken to accomplish this task. First, visual features with multi-instance representation are extracted to characterize the image due to the merits of bag-of-words representations. And a new distance function is designed for bags, which measures the relationship between bags more precisely. Second, the idea of multi-view learning is implemented since multiple views encourage the classifier to be more consistent and accurate. Last but not least, the metric learning technique is explored by optimizing the joint conditional probability to pursue view-dependent metrics and the importance weights of the newly-designed distance in multi-view scenario. Therefore, we propose a multi-view multi-instance metric learning method named MVMIML for image classification, which integrates the advantages of the multi-view multi-instance representation and metric learning into a unified framework. To solve MVMIML, we adopt the alternate iteration optimization algorithm and analyze the corresponding computational complexity. Numerical experiments verify the advantages of the new distance function and the effectiveness of MVMIML.