CN104537633B - A kind of method that utilization image fusion technology eliminates the anti-shadow of image - Google Patents

Abstract

The invention discloses a kind of method that utilization image fusion technology eliminates the anti-shadow of image, comprise the following steps：S1：Two images are gathered respectively under conditions of flash of light and non-flash labeled as flashlight images I_FWith non-flash image I_R；S2：By non-flash image I_RWith flashlight images I_FMark is non-flash image after alignment；S3：Position flashlight images I_FThe position of middle hot spot simultaneously calculates spot diameter；S4：To be alignd non-flash imageTone reversal is carried out, makes itself and flashlight images I_FTone be consistent；S5：According to one Gaussian mask of facula position and its attribute value generation, using the Gaussian mask by the alignment non-flash image after tone reversalWith flashlight images I_FSmooth is merged the image I for finally being eliminated anti-shadow.The present invention has that algorithm is easy and effective, the features such as require low to photographer and be easy to practical application.

Description

A kind of method that utilization image fusion technology eliminates the anti-shadow of image

Technical field

The invention belongs to technical field of image processing, and in particular to a kind of utilization image fusion technology eliminates the anti-shadow of image Method.

Background technology

When photographed scene is removed through transparent medium, the image of acquisition often shows as the mixed of background layer and reflecting layer Superposition is closed, background layer represents captured target scene, and reflecting layer is due to that light is produced in the reflex of media surface Some unknown scenes reflected image, this causes captured picture quality degradation.Such case is in daily life It can be frequently seen, such as the jewelry that glass protection is used in the fashionable dress shown in the show window of market, sales counter be shot, also in museum Calligraphy and painting historical relic of displaying etc..

At present, the method for eliminating the anti-shadow of image can be divided into by two major classes, the first kind according to the quantity of algorithm input picture Algorithm only requires input piece image, such as Levin and Weiss proposed a kind of anti-shadow removing method in 2004, and this method will User is asked first to mark out the marginal information for belonging to background layer and belonging to reflecting layer, then by the sparse of markup information and image gradient Prior information inputs an Optimized model Optimization Solution and obtains anti-shadow elimination result.Because this method needs user mutual, and And mark image edge information was not only time-consuming but also uninteresting, therefore it is not easy to application.Then, Levin proposes that a kind of automatic elimination is anti-again The algorithm penetrated, the algorithm constructs a natural image database first, is then searched using the thought of machine learning in image library Rope optimization background layer and the optimum combination in reflecting layer, so that the image for the reflection that is eliminated.But this based on sample Learning method is stronger to the dependence of database, and can also be failed when including more complicated structural information in image.The Two class methods require input multiple image, because multiple image can provide more useful informations for the solution of problem. N.Kong et al. proposed a kind of method using polarization in 2012, and polariscope is added before camera lens, inclined by rotation Light microscopic head gathers multiple image, and the mixed proportion that there is the image so gathered a feature to be exactly background layer and reflecting layer can be sent out Changing, the image for the reflection that can be eliminated based on this feature with optimization method.A.Agrawal et al. was proposed in 2005 Method in require gathering two images without using flash lamp and in the case of using flash lamp respectively, then using gradient throw The method of shadow eliminates reflection.This class method have certain professional technique requirement to shooting and meanwhile also require camera shooting when Time must be fixed, being capable of close alignment with the image for ensureing collection.But these requirements obviously increase for amateur photographers The difficulty shot is added, operation has been not easy in practice.

The content of the invention

It is an object of the invention to provide a kind of method that utilization image fusion technology eliminates the anti-shadow of image, this method needs Two images are shot respectively in the case where opening flash lamp and closing flash lamp.As shown in Fig. 2 opening taking photos by using flashlights image Anti- shadow can effectively be suppressed, can by hot spot " pollution ", although and close the image of taking photos by using flashlights will not be by light Spot " pollution ", but there is anti-shadow, it is possible to consideration is merged two images to eliminate anti-shadow and hot spot.

To achieve these goals, the present invention is adopted the following technical scheme that：One kind eliminates image using image fusion technology The method of anti-shadow, it is characterised in that comprise the following steps：

S1：Two images are gathered respectively under conditions of flash of light and non-flash labeled as flashlight images I_FAnd non-flash image I_R；

S2：By non-flash image I_RWith flashlight images I_FMark is non-flash image after alignment

S3：Position flashlight images I_FThe position of middle hot spot simultaneously calculates spot diameter；

S4：To be alignd non-flash imageTone reversal is carried out, makes itself and flashlight images I_FTone be consistent；

S5：According to one Gaussian mask of facula position and its attribute value generation, using the Gaussian mask by after tone reversal Alignment non-flash imageWith flashlight images I_FSmooth is merged the image I for finally being eliminated anti-shadow.

Wherein, in the step S2, flashlight images I is calculated respectively_FWith non-flash image I_RORB features, Ran Houyong PROSAC algorithms calculate the homography matrix obtained between two images, according to homography matrix to non-flash image I_RThrown Shadow is obtained and flashlight images I_FThe alignment non-flash image of alignmentIt is concretely comprised the following steps：

S2-1：Flashlight images I is calculated respectively_FWith non-flash image I_RFAST characteristic points with directional information, Ran Houyong BRIEF descriptors obtain ORB characteristic vectors to feature point description；

S2-2：Using PROSAC algorithms to flashlight images I_FWith non-flash image I_RORB characteristic vectors matched, count Calculation obtains homography matrix F；

S2-3：According to homography matrix F to non-flash image I_RRe-projection is obtained and flashlight images I_FThe alignment of alignment is non-to dodge Light image

In the step S3, flashlight images I_FThe positioning of middle hot spot and the calculating process of its property value are：

S3-1：Use Gaussian ProfileModel flashlight images I_FThe Luminance Distribution of middle hot spot, it is raw Into a width and highly equal Gaussian template, wherein u and v represent the row coordinate and row coordinate of pixel in Gaussian template respectively Position, σ span is flashlight images I_F0.01-0.05 times of width, the width of Gaussian template is set to 10* σ+1；

S3-2：With the Gaussian template and flashlight images I_FDo convolution algorithm, the maximum P of response_flashAs in hot spot Heart position, wherein

S3-3：Because the brightness of hot spot in the picture is very high, therefore image threshold is first obtained into bianry image, then Connection operation is carried out to bianry image and obtains multiple connected domains；

S3-4：Finally select the connected domain comprising spot center position and calculate its diameter as the effective straight of required hot spot Footpath d.

In the step S4, due to that can have colour-difference between image when being imaged to Same Scene under different illumination conditions It is different, in order that the image finally merged seems more natural, it is necessary to the non-flash image that will alignTone enter line translation and make it With flashlight images I_FTone be consistent.For alignment non-flash imageIn some given pixel x₀, after it is converted Color valueFor flashlight images I_FIn own the weighted sum of " adjacent " pixel color values, its operational formula is：

Wherein, C_F(x) it is flashlight images I_FMiddle pixel x color value, weights L (x, x₀) it is measurement pixel x and x₀" adjacent " The likelihood value of property, it meets condition

Pixel x and x₀The power of " adjacent " property is relevant with three factors：First be color similitude, use L_color(x, x₀) represent；Second be distance close degree, use L_dis(x,x₀) represent；3rd is far and near journeys of the pixel x apart from hot spot Degree, uses L_flash(x) represent.

L(x,x₀)=L_color(x,x₀)×L_dis(x,x₀)×L_flash(x)

Wherein

Wherein, | | | | represent 2- norms, C_R(x₀) represent conversion preceding pixel x₀Color value, α_colorIt is control parameter.Should Color difference between likelihood value and two pixels is inversely proportional, and color is more similar, and the value is bigger.

Wherein, P (x) represents pixel x coordinate position, α_disIt is control parameter.Distance between the likelihood value and two pixels It is inversely proportional, distance more close values are bigger.Consider distance factor to apart from x₀Near pixel, which distributes higher weights, can reduce image In the influence that is brought due to inhomogeneous illumination.

Wherein, P_flashThe center of required hot spot, α before expression_flashIt is control parameter.Because hot spot is to need to disappear Remove, so the point nearer apart from spot center position more needs to suppress, give these relatively low weights of point distribution, and to apart from light The higher weights of the point distribution of spot center farther out.

In the step S5, in order that the result of fusion seems more natural, generated using the hot spot attribute calculated before One Gaussian mask M, its center is spot center position, and σ values are the effective diameter of hot spot.Image I after being merged is

The method proposed by the present invention for eliminating the anti-shadow of image using flash of light and the fusion of non-flash image pair, with algorithm It is easy and effective, the features such as require low to photographer and be easy to practical application.

Brief description of the drawings

Fig. 1 is the flow chart that the present invention eliminates the anti-image method of image, and Fig. 2 is shot respectively under non-flash and flash conditions Non-flash image I_RWith flashlight images I_F, Fig. 3 is to non-flash image I_RBy the image alignd and tone reversal is obtained, Fig. 4 It is the Gaussian mask M used by the image co-registration and image I after fusion.

Embodiment

The above to the present invention is described in further details by the following examples, but this should not be interpreted as to this The scope for inventing above-mentioned theme is only limitted to following embodiment, and all technologies realized based on the above of the present invention belong to this hair Bright scope.

It is a kind of processing of one embodiment of the anti-shadow removing method of utilization image fusion technology of the present invention as shown in Figure 1 FB(flow block), comprises the following steps：

Step S1, inputs the two images gathered respectively under the conditions of flash of light and non-flash and is labeled as flashlight images I_FWith it is non- Flashlight images I_R, as shown in Figure 2.

Step S2, by non-flash image I_RWith flashlight images I_FNon-flash image of aliging is obtained after alignmentSuch as Fig. 3 (a) institutes Show.

The method of image alignment is using first then the feature of detection two images solves list to feature progress matching respectively Answering property matrix F, then wherein piece image is projected again according to matrix F can just make itself and another width image alignment.Wherein Feature can use existing various features to detect and description method, such as SHIFT features, SURF features, ORB features, this implementation ORB features are used in example.Homography matrix, which is solved, can use existing a variety of matching algorithms, and such as RANSAC algorithms, PROSAC is calculated PROSAC algorithms are used in method, the present embodiment.

It is further to note that the image alignment in the step both can be by non-flash image I_RWith flashlight images I_F Alignment or by flashlight images I_FWith non-flash image I_RAlignment, the two is substantially the same, and result will not be produced Substantial influence.The present embodiment is by non-flash image I_RWith flashlight images I_FAlign to illustrate.

Step S3, positioning flashlight images I_FThe position of middle hot spot and the diameter for calculating hot spot.

The present embodiment uses Gaussian ProfileModel flashlight images I_FThe brightness of middle hot spot point Cloth, for location spot position, one width of generation and highly equal Gaussian template, wherein u and v represent Gaussian template respectively The row coordinate and row coordinate position of middle pixel, σ span is flashlight images I_F0.01-0.05 times of width, Gaussian template Width be set to 10* σ+1, preferable result can be obtained, then with the Gaussian template and flashlight images I_FDo convolution algorithm, Its maximum P responded_flashThe as center of hot spot

The method for calculating spot diameter is that image threshold first is obtained into bianry image, and then bianry image is connected Operation obtains some connected domains.The connected domain comprising spot center position is finally selected as spot area, the straight of the region is asked Footpath as required hot spot effective diameter d.

Step S4, by non-flash image of aligingTone reversal is carried out, makes itself and flashlight images I_FTone is consistent, and is become Shown in image such as Fig. 3 (b) after changing.

For alignment non-flash imageIn some given pixel x₀, the color value after its conversionFor flashlight view As I_FIn own " adjacent " pixel color values weighted sum,

Wherein, C_F(x) it is flashlight images I_FMiddle pixel x color value, weights L (x, x₀) it is measurement pixel x and x₀" adjacent " The likelihood value of property, it meets condition

Pixel x and x₀The power of " adjacent " property is relevant with three factors：First be color similitude, use L_color(x, x₀) represent；Second be distance close degree, use L_dis(x,x₀) represent；3rd is far and near journeys of the pixel x apart from hot spot Degree, uses L_flash(x) represent.

L(x,x₀)=L_color(x,x₀)×L_dis(x,x₀)×L_flash(x)

Wherein

Wherein, | | | | represent 2- norms, C_R(x₀) represent conversion preceding pixel x₀Color value, α_colorIt is control parameter, takes It is worth for 0.1.Color difference between the likelihood value and two pixels is inversely proportional, and color is more similar, and the value is bigger.

Wherein, P (x) represents pixel x coordinate position, P (x₀) represent conversion preceding pixel x₀Coordinate position, α_disIt is control Parameter, value is the inverse of picture traverse.Distance between the likelihood value and two pixels is inversely proportional, and distance more close values are bigger.Examine Distance factor is considered to apart from x₀Near pixel, which distributes higher weights, can reduce the shadow brought in image due to inhomogeneous illumination Ring.

Wherein, P (x) represents pixel x coordinate position, P_flashThe center of required hot spot, α before expression_flashIt is control Parameter processed, value is 1/d.Because hot spot is to need to eliminate, the point nearer apart from spot center position more needs to suppress, The weights relatively low to these point distribution, and give longer-distance point distribution higher weights.

Step S5, according to one Gaussian mask M of facula position and its attribute value generation, is become tone using the Gaussian mask Alignment non-flash image after changingWith flashlight images I_FSmooth is merged the image I for finally being eliminated anti-shadow.

Utilize formulaGeneration one and the Gaussian mask M of input picture formed objects, such as Shown in Fig. 4 (a).Its center is P_flash, σ values are the effective diameter d of hot spot.Image I after fusion is

As shown in Fig. 4 (b), the image after fusion has effectively eliminated anti-shadow and has inhibited hot spot in flashlight images " pollution ".

The technological thought of above example only to illustrate the invention, it is impossible to which protection scope of the present invention is limited with this, it is every According to technological thought proposed by the present invention, any change done on the basis of technical scheme each falls within the scope of the present invention Within.

Claims (6)

1. a kind of method that utilization image fusion technology eliminates the anti-shadow of image, it is characterised in that comprise the following steps：

S1：Two images are gathered respectively under conditions of flash of light and non-flash labeled as flashlight images I_FWith non-flash image I_R；

S2：By non-flash image I_RWith flashlight images I_FMark is non-flash image after alignment

S3：Position flashlight images I_FThe position of middle hot spot simultaneously calculates spot diameter；

S4：To be alignd non-flash imageTone reversal is carried out, makes itself and flashlight images I_FTone be consistent, after conversion Color image pixel valueCalculated and obtained by equation below：

${\hat{C}}_R\left(x_0\right)=\underset{x\∈I_F}{\mathrm{\Σ}}L(x,x_0)\×C_F\left(x\right)$

Wherein, x₀Represent alignment non-flash imageIn some given pixel, C_F(x) it is flashlight images I_FMiddle pixel x color Value, weights L (x, x₀) it is measurement pixel x and x₀The likelihood value of " adjacent " property, it meets conditionAnd can be with Calculated and obtained by equation below：

L(x,x₀)=L_color(x,x₀)×L_dis(x,x₀)×L_flash(x)

Wherein, L_color(x,x₀) represent pixel x and x₀The similitude of color, L_dis(x,x₀) represent pixel x and x₀Distance it is close Degree, L_flash(x) represent pixel x apart from spot center position P_flashHow far；

S5：According to one Gaussian mask of facula position and its attribute value generation, using the Gaussian mask by pair after tone reversal Neat non-flash imageWith flashlight images I_FSmooth is merged the image I for finally being eliminated anti-shadow,

In the step S3, flashlight images I_FThe positioning of middle hot spot and the calculating process of its property value are：

S3-1：Use Gaussian ProfileModel flashlight images I_FThe Luminance Distribution of middle hot spot, generation one Individual width and highly equal Gaussian template, wherein u and v represent the row coordinate and row coordinate bit of pixel in Gaussian template respectively Put, σ span is flashlight images I_F0.01-0.05 times of width, the width of Gaussian template is set to 10* σ+1；

S3-2：With the Gaussian template and flashlight images I_FDo convolution algorithm, the maximum P of response_flashThe as centre bit of hot spot Put, wherein

S3-3：Image threshold is obtained into bianry image, connection operation is then carried out to bianry image obtains multiple connected domains；

S3-4：Finally select the connected domain comprising spot center position and calculate effective diameter d of its diameter as required hot spot.

2. the method that utilization image fusion technology according to claim 1 eliminates the anti-shadow of image, it is characterised in that：The step In rapid S2, flashlight images I is calculated respectively_FWith non-flash image I_RORB features, then with PROSAC algorithms calculate obtain two width Homography matrix between image, according to homography matrix to non-flash image I_RProgress, which is projected, to be obtained and flashlight images I_FAlignment Alignment non-flash imageIt is concretely comprised the following steps：

S2-1：Flashlight images I is calculated respectively_FWith non-flash image I_RFAST characteristic points with directional information, are then retouched with BRIEF State symbol and ORB characteristic vectors are obtained to feature point description；

S2-2：Using PROSAC algorithms to flashlight images I_FWith non-flash image I_RORB characteristic vectors matched, calculate To homography matrix F；

S2-3：According to homography matrix F to non-flash image I_RRe-projection is obtained and flashlight images I_FThe alignment non-flash figure of alignment Picture

3. the method that utilization image fusion technology according to claim 1 eliminates the anti-shadow of image, it is characterised in that：The step In rapid S4, color similarity likelihood value L_color(x,x₀) obtained by equation below calculating：

$L_{color}(x,x_0)=e^{-{\mathrm{\α}}_{color}\×\left|\right|C_F\left(x\right)-C_R\left(x_0\right)\left|\right|}$

Wherein, | | | | represent 2- norms, C_R(x₀) represent conversion preceding pixel x₀Color value, α_colorIt is control parameter, value is 0.1。

4. the method that utilization image fusion technology according to claim 1 eliminates the anti-shadow of image, it is characterised in that：The step In rapid S4, the close degree likelihood value L of location of pixels_dis(x,x₀) obtained by equation below calculating：

$L_{dis}(x,x_0)=e^{-{\mathrm{\α}}_{dis}\×\left|\right|P\left(x\right)-P\left(x_0\right)\left|\right|}$

Wherein, P (x) represents pixel x coordinate position, P (x₀) represent conversion preceding pixel x₀Position, α_disIt is control parameter, takes It is worth for flashlight images I_FThe inverse of width.

5. the method that utilization image fusion technology according to claim 1 eliminates the anti-shadow of image, it is characterised in that：The step In rapid S4, pixel x is apart from spot center position P_flashHow far likelihood value L_flash(x) calculated and obtained by equation below：

$L_{flash}\left(x\right)=1-e^{-{\mathrm{\α}}_{flash}\×\left|\right|P\left(x\right)-P_{flash}\left|\right|}$

Wherein, P (x) represents pixel x coordinate position, α_flashIt is control parameter, value is 1/d.

6. the method that utilization image fusion technology according to claim 1 eliminates the anti-shadow of image, it is characterised in that：The step In rapid S5, image co-registration step includes：

S5-1, utilizes formulaGeneration one and the Gaussian mask M of input picture formed objects, Its center is in P_flash, σ values are the effective diameter d of hot spot；

S5-2, by equation below by the alignment non-flash image after tone reversalWith flashlight images I_FMerged most The image I of anti-shadow is eliminated eventually：

$I=M\×\stackrel{\‾}{I_R}+(1-M)\×I_F.$