Hybrid Learning Framework for Large-Scale Web Image Annotation and Localization

In this paper, we describe the details of our participation in the ImageCLEF 2015 Scalable Image Annotation task. The task is to annotate and localize different concepts depicted in images. We propose a hybrid learning framework to solve the scalable annotation task, in which the supervised methods given limited annotated images and the searchbased solutions on the whole dataset are explored jointly. We adopt a two-stage solution to first annotate images with possible concepts and then localize the concepts in the images. For the first stage, we adopt the classification model to get the class-predictions of each image. To overcome the overfitting problem of the trained classifier with limited labelled data, we use a search-based approach to annotate an image by mining the textual information of its similar neighbors, which are similar on both visual appearance and semantics. We combine the results of classification and the search-based solution to obtain the annotations of each image. For the second stage, we train a concept localization model based on the architecture of Fast R-CNN, and output the top-k predicted regions for each concept obtained in the first stage. Meanwhile, localization by search is adopted, which works well for the concepts without obvious objects. The final result is achieved by combing the two kinds of localization results. The submitted runs of our team achieved the second place among the different teams. This shows the outperformance of the proposed hybrid two-stage learning framework for the scalable annotation task.