KDEVIR at ImageCLEF 2013 Image Annotation Subtask

The explosive growth of image data on the web leads to the research and development of visual information retrieval systems. However, these visual contents do not allow user to query images using semantic meanings. To resolve this problem, automatically annotating images with a list of semantic concepts is an essential and beneficial task. In this paper, we describe our approach for annotating images with controlled semantic concepts, which is a scalable concept image annotation subtask in the Photo Annotation and Retrieval task of the ImageCLEF 2013. We label training images with semantic concepts. After that, given a test image, the most k similar images are retrieved from the training image set. And finally, we extract and aggregate the concepts of the k matched training images, and choose the top n concepts as annotation. In our proposed method, the textual concepts of the training images are weighted by introducing BM25. Then, we utilizes some combination of visual features vectors, which are constructed from global descriptor such as color histogram, gist as well as local descriptor including SIFT and some variations of SIFT. The visual feature vectors are used to measure the similarity between two images by employing cosine similarity or inverse distance similarity (IDsim) that we introduce here. For a given test image, we find the k-nearest neighbors (kNN ) from the training image set based on the image similarity values. Furthermore, we aggregate the concepts of the kNN images, and choose top n concepts as annotation. We evaluate our methods by estimating F -measure and mean average precision (MAP). The result turns out that our system achieves the average performance in this subtask.