Improving OCR Performance With Word Image Equivalence

OCR is an error prone process when input images are degraded Most current OCR techniques use linguistic information such as character n grams or dictionaries to postprocess character recognition results These methods essentially discard the input image after the character recognition is complete This paper proposes a new technique for improving the performance of an OCR system that uses information about equivalent word images inside a document Words that are repeated inside a document are grouped into clusters by an image matching algorithm The decisions of an OCR algorithm about the identities of those words are used to generate a common recognition result for each of the original word images This technique thus combines information from the document image word image clusters with recognition results to correct errors made by OCR systems on di erent instances of the same word Experimental results are presented that show about of the words in a document are repeated two or more times A clustering algorithm is able to reliable locate a large percentage of these words in the presence of noise Experiments on images degraded with uniform noise show that the correct rate of a commercial OCR system can be improved from to on the words in those clusters An error analysis is given that shows with further development correct rates in the range are achievable Fourth Symposium on Document Analysis and Information Retrieval, Las Vegas, NV, April 24-26, 1995, pp. 177-190.