Image retrieval under very noisy annotations

In recent years, a significant number of tagged images uploaded onto image sharing sites has enabled us to create high-performance image recognition models. However, there are many inaccurate image tags on the Internet, and it is very laborious to investigate the percentage of tags that are incorrect. In this paper, we propose a new method for creating an image recognition model that can be used even when the image data set includes many incorrect tags. Our method has two superior features. First, our method automatically measures the reliability of annotations and does not require any parameter adjustment for the percentage of error tags. This is a very important feature because we usually do not know how many errors are included in the database, especially in actual Internet environments. Second, our method iterates the error modification process. It begins with the modification of simple and obvious errors, gradually deals with much more difficult errors, and finally creates the high-performance recognition model with refined annotations. Using an object recognition image database with many annotation errors, our experiments showed that the proposed method successfully improved the image retrieval performance in approximately 90 percent of the image object categories.