Needs for Estimating Page Quality

Estimating page quality for any given image would be beneficial for several applications:

• Controlling adaptive image processing for OCR. The existence of a way to automatically evaluate the quality of any given image would be essential for an adaptive image-enhancement algorithm. The algorithm would iteratively produce an image to be graded by the page quality estimator, which in turn would feedback the noise type or the degree of noise present in the image to the adaptive algorithm to generate the next (better) iteration of the image.

• Adaptive OCR algorithms. An image quality estimator would be essential to the operation of adaptive OCR algorithms since it could set the parameters for the OCR engine according to the quality of the page that it is about to process (Figure 1.3).
  

  

  Figure 1.3: Adaptive OCR Algorithms Architecture

• Reducing rekeying costs. As will be shown in this work, page/image quality is a direct cause of OCR errors. Therefore, estimating page quality can also provide an estimation of OCR accuracy. The minimum acceptable OCR accuracy for large-scale OCR operations is in the range of 95%-98% [4]. Correcting the errors on a page with less than 95%accuracy is more costly than retyping the page from scratch. A hypothetical ``OCR-accuracy estimator'' would act as a filter, classifying pages and filtering out those that would be better off rekeyed manually. In large-scale OCR environments, such a filter would represent substantial cost savings, since often the whole process is automated and the cost of manually rekeying a page after it has been processed implies disrupting the normal flow of the entire system.