2011 International Conference on Document Analysis and Recognition (2011)
Sept. 18, 2011 to Sept. 21, 2011
This article presents a new method to index document images. This work is done in an industrial context where thousands of document images are daily digitized, these images have to be sorted in different classes like payroll, various bills, information letters. We propose a software method which aims to accelerate this task. Usually, the number of document classes is a priori unknown. In this paper, we propose an automatic estimation of this class number. According to this class number, we use a clustering algorithm in order to group document images. After this step, we propose an assisted classification tool based on content based image retrieval method (CBIR). For each cluster, a reference image is automatically selected then considering a similarity measure, the other images are sorted and shown to the user. By interacting with the process, the user can reject wrong images. The user feedback is automatically taken into account to enhance the similarity measure by weighting each feature. The first tests show that, on average, databases are indexed 3 times faster with our assisted classification method than with a standard manual classification process.
document clustering, document retrieval, feature selection, relevance feedback, industrial application
O. Augereau, J. Domenger and N. Journet, "Document Images Indexing with Relevance Feedback: An Application to Industrial Context," 2011 International Conference on Document Analysis and Recognition(ICDAR), Beijing, China, 2011, pp. 1190-1194.