Statistical image differences, degradation features, and character distance metrics

Abstract

Abstract. Document image quality is degraded through processes such as scanning, printing, and photocopying. The resulting bilevel image degradations can be categorized based either on observable degradation features or on degradation model parameters. The image degradation features can be related mathematically to model parameters. In this paper we statistically compare pairs of populations of degraded character images created with different model parameters. The probability that the character populations were degraded by the same model parameters correlates with the relationship between observable degradation features and the model parameters. Two metrics of character difference are used: Hamming distance and moment feature distance. Knowledge of under which conditions characters will be similar and when they will be different can influence the choice of parameters for future experiments.

<title>Partitioning of the degradation space for OCR training</title>

“…When the division was made based on edge spread, the recognition rates of the problem improved even more to 97.6%. These improved recognition rates are consistent with observations that characters that had the same edge spread both appeared similar [4] and were statistically similar [5]. What remained un-clear was how to choose the values of the edge degradation with which to divide the degradation space.…”

“…This results in the centroids of the cluster moving upward and thus the degradation boundaries "rising" to regions of higher thresholds, or to be more precise, more negative edge spread. Also the thicker strokes mean that the assumption that the strokes are affected by the edge spread independently is more valid so the cluster boundaries now have a greater resemblance to the theoretical edge spread lines predicted in [5].…”

<title>Correlating degradation models and image quality metrics</title>

“…The amount that an edge is displaced, b c [3], and the amount a black or white comer is eroded, db and d w [4,16]. Later work showed that the images are more affected by the edge spread [2] than they are by the corner erosion. It was also shown that OCR accuracy could be improved by grouping characters with common edge spreads [5].…”

Human Image Preference and Document Degradation Models

“…These were chosen so there would be a character with many sharp corners and a character with no corners. These were also the characters used in [2]. The survey used samples with five different edge spread amounts and three different PSF widths for a total of 15 combinations of parameters for each character.…”

“…Two resulting effects of the degradation have been defined: the amount an edge is displaced, δ c , and the amount a black or white corner is eroded [3,4]. A statistical test was conducted in [2] to compare the similarity between groups of characters synthetically generated with parameters (w, Θ) varying over the parameter space. This test showed that the amount of variation in the characters correlated highly with the amount of edge spread,…”