CN104809733A - Ancient building wall polluted inscription character image edge extraction method - Google Patents

Abstract

The invention discloses an ancient building wall polluted inscription character image edge extraction method and belongs to the field of ancient building digitalized repairing. The method includes: performing total variation denoising and multiscale retina enhancement on a collected image; forming a shadow-shelter quasi invariant of the enhanced image, eliminating pseudo edges generated by pollution and background, using a gabor filter to performing space averaging on the enhanced image, establishing a color structure tensor, and combining with a Canny algorithm for edge extraction; processing an edge image through a mathematical morphology method and a neighborhood search algorithm, filtering sundry edges, and connecting character edges to acquire an inscription character edge image. By the method, pseudo edges and noncontinuous edges caused by pollution can be recognized and eliminated effectively, and inscription character edges can be extracted well. The method is mainly used for (not limited to) edge extraction of ancient building wall polluted inscription characters.

Description

A kind of ancient architecture wall is got dirty and is inscribed note character image edge extracting method

Technical field

The invention belongs to ancient architecture Digital repair field, be specially a kind of ancient architecture wall eliminating the phenomenons such as the pseudo-edge that produces in edge extracting process and discontinuous edge and get dirty topic note character image edge extracting method.

Background technology

Writing brush word on ancient architecture wall, carries a large amount of historical information, simultaneously the living specimen of Ye Shi China calligraphy, has very large aesthetic and historic survey to be worth.But due to long-term weathering, ink marks, by the pollution of mushroom, goes mouldy serious, the corrosion of wall self and aging qualitative change in addition, many writings cannot identification.Artificial entity reparation easily causes the destruction again of original writing while removing pollution, in recent years, edge extracting is as the important technology in image procossing and computer vision, great concern is obtained in ancient architecture Digital repair field, digital processing technology is relied on to carry out edge extracting to ancient architecture wall character image, contribute to the identification of word on the one hand, significant for its calligraphy stroke of research, on the other hand for text entity reparation provides strong guidance, ensure carrying out smoothly of repair.

In recent years, the edge extraction techniques for character image has based on mathematical morphology, based on transform domain with based on methods such as gradient operators.Method based on mathematical morphology can ensure the integrality at word edge, but not high for the identity of pollution and true edge in image, and pseudo-edge is more; Based on the method for transform domain, as wavelet transform modulus maxima method, there is good robustness to high frequency noise, but the character image under large area bulk is polluted, extraction effect is bad; Based on the method for gradient operator, as Canny operator, Sobel operator, edge extracting is more coherent, and details retains better, but still the pseudo-edge produced that cannot decontaminate.Said method is better for of reduced contamination, that background is single pictograph extraction effect, but when seriously polluted affect word marginal information time, the edge extracted usually contains more pseudo-edge and discontinuous edge.

Summary of the invention

The present invention is for solving the note contaminated impact of character image of ancient architecture wall topic and edge extracting is imperfect, the problems such as easy generation pseudo-edge and discontinuous edge, propose one and to get dirty topic note character image edge extracting method in conjunction with illumination model and morphologic ancient architecture wall.This method removes tiny noise by pre-service, outstanding dark areas details; The pseudo-edge brought is polluted in Canny operator filtering in conjunction with illumination model and improvement; To be mixed limit by morphology and regional connectivity process filtering, connect edge, ensure that integrality and the continuity at word edge.

The present invention adopts following technical scheme to realize: a kind of ancient architecture wall is got dirty and inscribed note character image edge extracting method, comprises the following steps:

S1: utilize digital camera collection to inscribe note character image, during shooting, camera lens is vertical with wall to be placed, and the resolution of collection image is at least more than 200dpi, size 2048 × 1536 pixel;

S2: full variation denoising is carried out, the random noise produced in filtering gatherer process to the image that S1 obtains, obtains denoising image;

S3: utilize multiple dimensioned retina algorithm to carry out enhancing process to denoising image, the text details information of outstanding dark areas, so that subsequent edges extracts;

S4: calculate the space differentiation of image that S3 obtains, and it is made projection process in color space direction, the edge that pollution abatement and shade produce, obtains only containing the shade-block accurate invariant at material edge;

S5: utilize the inner product of gabor wave filter to the accurate invariant of shade-block to carry out space average, protect word marginal information preferably while edge accumulation, then constructs shade-block accurate constant color tensor;

S6: the dominant eigenvalue λ calculating the tensor formula that S5 obtains ₁, with as the gradient magnitude of Canny algorithm, then carry out the edge image that maximum value suppresses and obtains word after dual threshold process;

S7: utilize Mathematical Morphology Method and nearly Neighborhood-region-search algorithm edge image to process, the tiny assorted limit in elimination edge image, and connect word discontinuous edge, ensures integrality and the continuity of strokes of characters, obtains final word edge extracting image;

S8: the word edge extracting Image Saving that S7 is obtained or output.

Above-mentioned a kind of ancient architecture wall is got dirty and is inscribed note character image edge extracting method, and described shade-blocking accurate constant color tensor structure carries out according to the following steps:

S51: structure gabor wave filter h _{ω, θ}(x, y), wherein, x, y are pixel coordinate, and θ is filter direction, and for suppressing edge fog, be taken as the equiluminous direction of each pixel of image, ω is modulating frequency, according to character image texture feature value;

S52: carry out space average by the inner product of gabor wave filter to the accurate invariant of shade-block, obtains shade-block accurate constant color tensor G _g, be calculated as follows formula:

$G_g={\left\{g_{\mathrm{tk}}\right\}}_{t,k=\mathrm{1,2}}=\left(\begin{array}{cc}\mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_y^c)\\ \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_y^c)\end{array}\right)$

Wherein, be shade-block accurate invariant.

Get equiluminous direction as the average direction of gabor filter space, can ensure that in diffusion process, image border can not obtain too many level and smooth and fuzzy, is beneficial to subsequent edges and extracts.

Above-mentioned a kind of ancient architecture wall is got dirty and is inscribed note character image edge extracting method, and described edge image Morphological scale-space carries out according to the following steps:

S71: utilize Mathematical Morphology Method edge image to process: first, each connected region of edge image marks, adjacent gray-scale value is each pixel composition marked region of 1, and adds up each marked region pixel number as this region area; Then set threshold value to filter marked region, elimination area is less than the connected domain of given threshold value; Finally, carry out morphology opening operation to image: C=H ο Q, wherein, H is edge image, and ο represents morphology opening operation, and Q is structural element.

Utilize morphological method to mark connected region, and carry out area and filter and the process of morphology opening operation, can the tiny assorted limit of filtering to greatest extent.

The present invention compared with prior art has the following advantages:

1. the present invention utilizes full variation denoise algorithm and multiple dimensioned retina enhancing algorithm to carry out pre-service to topic note character image; the marginal information of word is well protected while filtering random noise; and the dark areas text details of image can be given prominence to significantly, be beneficial to subsequent edges and extract.

2. in conjunction with illumination model and color tensor, the gradient modulus value item to Canny operator improves in the present invention, can distinguish the edge and word true edge that pollute and produce preferably, the pseudo-edge that pollution abatement produces.

3. the present invention is processed by Mathematical Morphology Method and neighbor search algorithm edge image, the tiny assorted limit of filtering image, and is communicated with discontinuous word edge, ensure that integrality and the continuity of strokes of characters.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is the topic note character image gathered;

Fig. 3 is full variation denoising image;

Fig. 4 is that retina strengthens image;

Fig. 5 is edge-detected image;

Fig. 6 is Morphological scale-space image;

Fig. 7 is edge extracting result figure;

Fig. 8 is 3rd level Edge extraction Contrast on effect, in figure, (a) is topic note character image, b () is Canny algorithm extraction effect figure, (c) is Sobel algorithm extraction effect figure, and (d) is the inventive method extraction effect figure;

Fig. 9 is the 4th grade of Edge extraction Contrast on effect, in figure, (a) is topic note character image, b () is Canny algorithm extraction effect figure, (c) is Sobel algorithm extraction effect figure, and (d) is the inventive method extraction effect figure;

Figure 10 is the 5th grade of Edge extraction Contrast on effect, in figure, (a) is topic note character image, b () is Canny algorithm extraction effect figure, (c) is Sobel algorithm extraction effect figure, and (d) is the inventive method extraction effect figure.

Embodiment

With reference to the process flow diagram of Fig. 1, for eight scape topic note words in Jinci Monastery museum of Shanxi Province, test, concrete implementation step is as follows:

S1: utilize digital camera collection to inscribe note character image, during shooting, camera lens is vertical with wall to be placed, and Fig. 2 is the sub-picture example gathered;

S2: utilize full variation denoising model to carry out denoising to the image that S1 obtains, according to feature of image, the full variation denoising model adopted is such as formula (1):

$E=\underset{u}{\arg \min }\mathrm{VTV}\left(u\right)+\frac{\mathrm{\λ}}{2}{\left|\right|I-u\left|\right|}_2^2---\left(1\right)$

Wherein, E represents energy functional, and I is for gathering image, and u is denoising image, and λ is regularization parameter, and span is [0,1], and this example gets λ=0.05, for the full variation item of colour, be used for restraint speckle; for regular terms, be used for Protect edge information, obtain denoising image u by solving energy functional minimum value, Fig. 3 is denoising image;

S3: carry out multiple dimensioned retina to denoising image and strengthen process, with outstanding dark areas text details information, for raising the efficiency, each amount is converted to logarithmic form and carries out interative computation, such as formula (2) by embodiment:

log[f(x,y)]＝log[f _i-1(x,y)]+ω ₀*(log[u _i(x,y)]-log(l _i(x,y))) (2)

Wherein, u _i(x, y) is the denoising image of the i-th yardstick, l _i(x, y) represents luminance component, and obtain by carrying out Gaussian Blur to image, this example makes l _i=G _σ* u _i, G _σfor Gaussian function, σ is gaussian kernel, and span is (0, ∞), and this example gets σ=3, ω ₀for weight coefficient, span is [0,1], and this example gets equal weight, and f (x, y) represents reflecting component, namely strengthens the image after process, and Fig. 4 is for strengthening image;

S4: calculate the space differentiation of image that S3 obtains, and it is made projection process in color space direction, the edge that pollution abatement and shade produce, obtains only containing the shade-block accurate invariant at material edge;

Concrete steps are as follows:

S41: the space differentiation asking for image in S3: f _x=(R _x, G _x, B _x), R, G, B are image red, green, blue passage;

S42: space differentiation is projected to the shade of image-block direction on, obtain shade-block accurate invariant wherein, for the vector of unit length in color of image direction, computing formula is: $\hat{f}=1/\sqrt{R^2+G^2+B^2}{(R,G,B)}^T;$

S5: utilize the inner product of gabor wave filter to the accurate invariant of shade-block to carry out space average, protect word marginal information preferably while edge accumulation, then constructs shade-block accurate constant color tensor;

S51: construct two-dimentional gabor wave filter h _{ω, θ}(x, y), wherein, x, y are pixel coordinate, and θ is filter direction, and for suppressing edge fog, be taken as the equiluminous direction of each pixel of image, ω is modulating frequency, and this example, according to character image texture feature, selects ω=0.3;

S52: carry out space average by the inner product of gabor wave filter to the accurate invariant of shade-block, obtains shade-block accurate constant color tensor G _g, calculate such as formula (3):

$G_g={\left\{g_{\mathrm{tk}}\right\}}_{t,k=\mathrm{1,2}}=\left(\begin{array}{cc}\mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_y^c)\\ \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_y^c)\end{array}\right)---\left(3\right)$

S6: the dominant eigenvalue λ calculating the color tensor that S5 obtains according to formula (4) ₁, will gradient magnitude as Canny operator carries out maximum value to be suppressed and dual threshold process, and obtain edge image H, Fig. 5 is edge image;

${\mathrm{\λ}}_1=\frac{1}{2}(g_{11}+g_{22}+\sqrt{{(g_{11}-g_{22})}^2+{(2\·g_{12})}^2})---\left(4\right)$

S7: utilize Mathematical Morphology Method and nearly Neighborhood-region-search algorithm edge image to process, the tiny assorted limit in filtering image, and be communicated with word discontinuous edge, ensures integrality and the continuity at character image edge;

S71: utilize Mathematical Morphology Method edge image to process, concrete steps are: first, the each connected region of edge image marks, and adjacent gray-scale value is each pixel composition marked region of 1, and adds up each marked region pixel number as this region area; Then set threshold value T to filter marked region, elimination area is less than the connected domain of T, and this example gets T=8; Finally, carry out morphology opening operation to image: C=H ο Q, wherein, H is edge image, and ο represents morphology opening operation, the structural element of Q to be radius be r, r span be [1, ∞), this example gets r=5, and Fig. 6 is image after Morphological scale-space;

S72: utilize nearly Neighborhood-region-search algorithm to carry out regional connectivity to image in S71, concrete steps are: centered by each pixel, investigate the gray-scale value situation of its eight neighborhood pixel, if there is the pixel that gray-scale value is non-vanishing in its neighborhood, then think that two pixels are a pair discontinuous edge point, both connected, iteration investigates each pixel, obtain final edge extracting image, Fig. 7 is final edge extracting figure;

S8: the word edge extracting result obtained by S7 is preserved or exported.

For fully verifying the efficiency and applicability of the inventive method, according to the pollution level of word, the image of collection is divided into 5 grades, as shown in table 1, the present invention mainly carries out experimental study to rear 3 grades of character images.

Table 1

Rank	Title	Feature
			1st grade	Pollution-free	Character area is pollution-free, does not affect marginal information
2nd grade	Light pollution	Word goes mouldy comparatively light, does not affect marginal information
			3rd level	General pollution	Segment word goes mouldy heavier, affects edge extracting
4th grade	Heavily contaminated	Major part word is contaminated, fuzzyly not easily recognizes
			5th grade	Full pollution	Font and pollution combine together, are difficult to identification

As seen from Figure 8, for 3rd level character image, the inventive method compares traditional C anny algorithm and Sobel algorithm has better extraction effect, well inhibits pseudo-edge and discontinuous edge, and word integrality is better; Fig. 9 can find out, for the 4th grade of character image, the inventive method extraction effect is remarkable, better to the robustness of pollution and noise; Figure 10 can find out, for the 5th grade of character image, the inventive method effectively can identify word and pollution of going mouldy, and edge extracting is continuous, complete, and extraction effect is better.

For quantitative test Edge extraction effect, the present invention edge quality EQ carries out objective evaluation to it: Ec is image fine edge number, is used for the quantity of pseudo-edge in characterizing edges image; Rm is largest connected territory area, is used for the integrality of target in characterizing edges image, and EQ is defined as the ratio of largest connected territory area Rm and fine edge number Ec.Table 2 is the index contrast of rear 3 grades of image zooming-out effects, can find out, compare Canny method and Sobel method, the Ec of the inventive method is less, Rm is comparatively large, and edge quality EQ is high, show its pseudo-edge remove and word edge integrality on have and comparatively significantly promote.

Table 2

Claims (3)

1. ancient architecture wall is got dirty and is inscribed a note character image edge extracting method, it is characterized in that comprising the following steps:

S1: fixing digital camera, makes camera lens remember that wall is vertical with topic and places, and collection topic note character image, gathers the resolution at least 200dpi of image, more than size 2048 × 1536 pixel;

S2: full variation denoising is carried out to the image that S1 obtains, the random noise produced in filtering image acquisition and transmitting procedure, obtains denoising image;

S3: utilize multiple dimensioned retina algorithm to carry out enhancing process to denoising image, the text details information of outstanding dark areas, so that subsequent edges extracts;

S4: calculate the space differentiation of image that S3 obtains, and it is made projection process in color space direction, the edge that pollution abatement and shade produce, obtains only containing the shade-block accurate invariant at material edge;

S5: utilize the inner product of gabor wave filter to the accurate invariant of shade-block to carry out space average, protect word marginal information preferably while edge accumulation, then constructs shade-block accurate constant color tensor;

S6: the dominant eigenvalue λ calculating color tensor in S5 ₁, will as the gradient modulus value in Canny algorithm, carry out maximum value and suppress and obtain edge image after dual threshold process;

S7: utilize Mathematical Morphology Method and nearly Neighborhood-region-search algorithm edge image to process, the tiny assorted limit in filtering image, and be communicated with word discontinuous edge, ensures integrality and the continuity of word edge image, obtains final word edge extracting image;

S8: the word edge extracting Image Saving that S7 is obtained or output.

2. a kind of ancient architecture wall according to claim 1 is got dirty topic note character image edge extracting method, it is characterized in that described in S5 shade-blocking accurate constant color tensor structure carries out according to the following steps:

S51: structure gabor wave filter h _{ω, θ}(x, y), wherein, x, y are pixel coordinate, and θ is filter direction, and for suppressing edge fog, be taken as the equiluminous direction of each pixel of image, ω is modulating frequency, according to character image texture feature value;

S52: carry out space average by the inner product of gabor wave filter to the accurate invariant of shade-block, obtains shade-block accurate constant color tensor G _g, be calculated as follows formula:

$G_g={\left\{g_{\mathrm{tk}}\right\}}_{t,k=\mathrm{1,2}}=\left(\begin{array}{cc}\mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_x^c\·S_y^c)\\ \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_x^c)& \mathrm{\Σ}{h}_{\mathrm{\ω},\mathrm{\θ}}(S_y^c\·S_y^c)\end{array}\right)$

Wherein, be shade-block accurate invariant.

3. a kind of ancient architecture wall according to claim 1 is got dirty and is inscribed note character image edge extracting method, it is characterized in that the edge image Morphological scale-space described in S7 carries out according to the following steps:

S71: utilize Mathematical Morphology Method edge image to process: first, each connected region of edge image marks, adjacent gray-scale value is each pixel composition marked region of 1, and adds up each marked region pixel number as this region area; Then set threshold value to filter marked region, elimination area is less than the connected domain of given threshold value; Finally, carry out morphology opening operation to image: C=H o Q, wherein, H is edge image, and o represents morphology opening operation, and Q is structural element.