Using morphology in document image processing

Abstract

ABSTRACT To improve the readability, image compression, and optical character recognition (OCR) system performance fortwo-tone (binary) text image data, we investigated morphological methods of image processing. We found them to be fastand effective not only with noisy text images but with relatively noise-free images as well. Using morphology, weimproved text image readability as judged in a blind test, increased compression ratio using CCITF Group 4, and reducedOCR error (cluster) rates in a commercial omnifont scanner. 1. INTRODUCTION Technology advances continue to drive down the costs of electronic data processing, storage, and networking whilethe costs of handling paper and microfilm are increasing. As a consequence, many government agencies and privateorganizations that process large quantities of paper or microfilm are seeking to reduce their processing costs throughconversion of their documents to electronic form. Ideally, these documents will be converted to character format ratherthan raster image format, thus allowing full text searching and more efficient use of storage and bandwidth. However,because the documents often contain text in fonts that are not known a priori, the process of conversion requires anomnifont optical character recognition (OCR) system. Unfortunately, image noise tends to slow the OCR process andincrease recognition errors. There is thus motivation to decrease the noise in the document images prior to OCRprocessing. Furthermore, even if the documents are only stored in raster image format, noise can significantly increase thestorage requirements as a consequence of the run-length based compression techniques usually employed.Although preprocessing of gray scale images can substantially reduce image noise, document images are oftenobtained from scanners which only provide bitonal (binary) output There are also large inventories of document imageswhich have already been thresholded (converted into binary images). The choice of image processing techniques is,therefore, often limited to those which operate on binary images.We have investigated morphological methods of binary image processing1 in an attempt to develop techniqueswhich are fast and effective in removing the noise that is often present in binary document images. We have found thatsuch methods can be effective in selectively removing certain types of image noise, and thereby, in improving imagecompression ratios and, to a limited degree, improving the performance of omnifont OCR systems.All document images used in this study were captured in a production mode scanning operation. The noise wastherefore naturally acquired (either during scanning or inherent in the paper document as caused by bleed-through andyellowing) and not artificially synthesized.The sections that follow will discuss morphological methods in binary image processing, our methodology, ourresults, an implementation of morphological filters, and our conclusions.