Handling imperfections for multimodal image annotation

摘要: This thesis deals with multimodal image annotation in the context of social media. We seek to take advantage textual (tags) and visual information order enhance performances. However, these tags are often noisy, overly personalized only a few them related semantic content image. In addition, when combining prediction scores from different classifiers learned on modalities, faces their imperfections (uncertainty, imprecision incompleteness). Consequently, we consider that is subject at two levels: representation decision. Inspired fusion theory, focus this defining, identifying handling imperfection aspects improve annotation.