In this experiment, we assume that colour of a paper with
a document is white and that the size of a paper in this paper
is known as A4. The size of a normal A4 paper is 210mm x
290mm, which are used in image warping step. In the initial
white region detection step, a fixed threshold, th in (2), is set
as 110. The resolution of captured images is 3072 X 2048.
Fig.7 shows examples of captured images and results. We
choose a desk with relatively white colour and document
with many characters to verity the robustness of the proposed
method as shown in Fig.7. Although Fig.7(a) has some white
texture on the clothe and Fig.7(b) has some line segments
in the document, we succeed in extracting the feature points
to estimate a homography. Fig.7(c) and Fig.7(d) show the
final captured documents from images. In Fig.8, the four
green points are detected feature points used to estimate a
homography, and the four red solid lines are the extracted
boundary of a paper from the image Fig.7(a).
The experimental results show that the proposed method
successfully removes text regions, which are outliers of finding
line features to eliminate perspective distortions. A document
could be captured from the edge image without text regions.
However, if there are some object with line segments in
other areas, not in a paper region, the proposed method does
not succeed to extract feature points in some cases. Even if
outliers from text regions in a paper are removed, the image
might have other outliers. This is the problem of selecting
the detected line features. The image without text region
has much less line segments than the one with text regions.
We eXpect that it could be improved in the future work.
Additionally, captured images sometimes need some other
sophiscated post processing and restoration, if it would be
used in other applications such as OCR. For example, the left
column of Fig.9 has some distortion due to curled paper,