Pyrros Koletsis
Image annotation is the task of assigning a class name or description to an unknown image. In this work, we propose SIA, a framework capable of automatically annotating images using information from ontologies in combination with low level image features (color and texture) which are extracted from raw image data. The method works for images of a particular domain. First, an ontology is constructed denoting characteristics of the various image classes in this domain. A set of low level image characteristics is also assigned to each class. Image annotation is then implemented as a retrieval process by comparing vectors of such low-level characteristics extracted from the input image and representative images of each class in the ontology respectively. A combined similarity measure is used between images. The relative importance of low-level features in this measure is determined using machine learning by decision trees. The result list of images are ranked in decreasing visual similarity. AVR(Average Retrieval Rank) is used as a metric to estimate the semantic category where the image is possible to belong to (ie. the unknown image is assigned a class which is computed by voting among the top ranked retrieved images from the ontology). The experimental results demonstrate that approximately 70% of the input images are correctly annotated (ie. the method identified its class correctly). Experiments and evaluations were realized on an image dataset consisting of images belong to 30 dog breeds (semantic categories), which were collected from the World Wide Web (WWW).
No comments:
Post a Comment