Edges are the features that are usually more salient and robust
than points for pattern recognition problems such as document
recognition and analysis. In [4], the text area were detected
by using a stoke width transform based method where
the stroke width was computed according to edges. To estimate
the PSF (Point Spread Function) of the image, Joshi [5]
proposed an edge prediction method which assumed the high
quality image has sharp edges and the blurry image’s edges
are smoothed. Inspired by these ideas and based on the observation
that the gradient of the ideal text edge changes rapidly
from the stroke to the background area while the gradient of
the degraded text edge changes smoothly, we use gradient of
the edge to measure the degradation.
Prior to the feature extraction, the edges of the document
image are detected using the Canny operator [6]. A gradient
search algorithm is applied to each edge pixel px;y to find
the pixel such that the gray difference between two pixels is
bigger than a predefined threshold with the shortest distance.
The searching algorithm is summarized in Fig. 2
The average distance of all edge pixels within a page is
calculated according to: