Abstract: On the issue that inscription text image of the ancient wall is contaminated leading to its edge
extraction is not complete, easy to produce false edge and discontinuous margin, this paper proposes a
novel method of edge extraction from polluted inscription text. Firstly, text image for total variation
denoising and multi-scale retinal algorithm shading to enhance preprocessing is used. Then construct a
shadow-blocking quasi-invariant model filtering out pollution and pseudo-edge generated by
background. Meanwhile, use Gabor filter to do spatial averaging and establish color structure tensor
combined with Canny algorithm to do edge detection. Finally, filter small area edge existing in the
image by morphological method and do the regional connectivity to obtain the edge inscription text
image. The results show that, compared to traditional Canny algorithm and morphological method, this
method effectively overcomes the false edge and discontinuous edge caused by dirt and background
which makes good extraction of edge inscription from the full text and a more objective indicators such
that edge quality has improved significantly.

Key words: inscription text, false edge, total variation, color structure tensor