CN104700381B - A kind of infrared and visible light image fusion method based on well-marked target - Google Patents
A kind of infrared and visible light image fusion method based on well-marked target Download PDFInfo
- Publication number
- CN104700381B CN104700381B CN201510111415.0A CN201510111415A CN104700381B CN 104700381 B CN104700381 B CN 104700381B CN 201510111415 A CN201510111415 A CN 201510111415A CN 104700381 B CN104700381 B CN 104700381B
- Authority
- CN
- China
- Prior art keywords
- image
- visible images
- infrared
- notable
- infrared image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Landscapes
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of infrared and visible light image fusion method based on well-marked target, includes the following steps:The infrared image and visible images for including several targets to giving Same Scene are established infrared and visible images Nonlinear Scale Space Theories and are indicated respectively;On the basis of image non-linear scale space indicates, infrared and visible images vision attention notable figures are calculated using visual attention computation model;On the basis of infrared and visible images vision attention notable figure, the well-marked target region in infrared and visible images is selected respectively using inhibition of return mechanism, and calculate all well-marked target regions in entire scene;Registration operation is carried out to infrared image and visible images, fusion treatment is carried out using Pixel-level blending algorithm to well-marked target region, to non-significant target area, fusion treatment is carried out using feature-based fusion algorithm;Synthesis result generates the infrared image and visual image fusion image of whole scene.
Description
Technical field
The present invention relates to a kind of multi-source image integration technology field, it is especially a kind of based on well-marked target it is infrared with it is visible
Light image multi-level Fusion method.
Background technology
Image co-registration (Image Fusion) refers to that multiresolution or multimedium image data are passed through spatial registration and figure
As message complementary sense generates the Comprehensive Analysis Technique of new image.Compared with single-sensor image, blending image can be to greatest extent
The information using each source images, improve resolution ratio and clarity, increase sensitivity, perceived distance and the essence of image object perception
Degree, anti-interference ability etc., to reduce the imperfection and uncertainty of target apperception, improve to the accuracy rate of target four and
Scene interpretation ability.The general flow of image co-registration is as shown in the figure.First, some pretreatment operations are carried out to several source images,
Such as filter out noise;Secondly, time-space relation is carried out to image, that is, mapped them into the same space-time coordinates, to ensure
The consistency of their space-time positions;Again, the image after registration is handled using corresponding method;Finally, according to certain
Fusion rule carry out fusion treatment obtain blending image.
Photoelectricity and infrared detection sensor are two kinds of most commonly seen load of the tactical reconnaissances platform such as early warning plane and unmanned plane.
Visual light imaging is in 0.4~1.0 μm of wave band, its main feature is that high resolution, good concealment can get abundant contrast, color
And shape information, but influenced by ambient light illumination, it cannot work at night, no camouflage recognition capability.Infrared imaging is in 8~14 μm of wave band
(long infrared band) and 3~5 μm (middle infrared bands), its main feature is that with the ability for centainly penetrating cigarette, mist, snow etc., concealment
Good, imaging resolution is higher, and detection range is generally in several kms between more than ten kms, but climate influences, operating distance compared with
Image is not sufficiently stable when remote.When each spot temperature of target itself changes greatly or thermal background emission characteristic is weaker, infrared figure
As comprising target or background detailed information it is less, and visible images then may include abundant detailed information;However, in light
Line is dark or has in cigarette, cloud, mist bad border, it is seen that light image is second-rate, and the target in infrared image is but still clear and legible.It will
Visible images carry out fusion treatment with infrared image, and ambient noise can be overcome to grasp mesh comprehensively to the influence of target acquisition
Target shape and minutia, and do not influenced by ambient lighting, it round-the-clock can use, to improving reconnaissance platforms detection and figure
System performance is studied and judged as intelligence analysis to be of great significance.
The key tactical that US Department of Defense formulates in different times in the works, has significant component of task to be related to multi-source figure
In terms of fusion." LANTIAN " gondola that good operational performance is played in the Gulf War be exactly one kind can by forward-looking infrared, swash
The image fusion system that the multiple sensors information superposition such as ligh-ranging, visible light camera is shown.American TI Company is in nineteen ninety-five
It is obtained from U.S.'s night vision and electronic sensor management board (NVESD) by infrared with three generations's LLL image fusion integrated design to first
Into the contract of helicopter (AHP) sensing system, signal processing function is completed by TMS32OC30DSP.2003, U.S. national defense
Portion proposes " the lateral fusion (Transformational about Horizontal Fusion) in Military transforming " white paper,
Focus on laterally fusion especially data fusion, it is therefore an objective to the lateral interoperability of strengthening system especially data fusion ability.
2006, general office of US Department of Defense (OSD) issued the Annual Technical bulleted list to be verified within the coming years, to push away
Move the fast development in these fields.
The work that the country is merged in multi-source image concentrates on algorithm research, including the use of DT-CWT transformation, multi-scale wavelet
The methods of sampling, region segmentation research multi-source image fusion.Existing systems technology is confined to the difference to Same Scene mostly
The simple superposition of the image pixel of phase or its different color channels accepts or rejects operation.
Invention content
Goal of the invention:The technical problem to be solved by the present invention is to be directed to existing infrared image and visual image fusion side
It is insufficient present in method, a kind of infrared and visible light image fusion method based on well-marked target is provided.
In order to solve the above-mentioned technical problem, the invention discloses a kind of, and the infrared and visible images based on well-marked target melt
Conjunction method, includes the following steps:
Step 1, the infrared image and visible images for including several (more than one) targets to giving Same Scene, respectively
Infrared and visible images multiscale spaces are established to indicate;
Step 2, it on the basis of image non-linear scale space indicates, is calculated using visual attention computation model infrared and can
The vision attention notable figure of light-exposed image.
Step 3, it on the basis of infrared and visible images vision attention notable figure, is selected respectively using inhibition of return mechanism
Go out the well-marked target region in infrared and visible images, and calculates all well-marked target regions in entire scene.
Step 4, registration operation is carried out to infrared image and visible images, well-marked target region is merged using Pixel-level
Algorithm carries out fusion treatment, and to non-significant target area, fusion treatment is carried out using feature-based fusion algorithm.
Step 5, comprehensive well-marked target region fusion results and non-significant region fusion results, generate the infrared figure of whole scene
Picture and visual image fusion image.
Step 1 of the present invention establishes infrared and visible images multiple dimensioned skies including the use of the method for Nonlinear Scale Space Theory
Between indicate;The multiscale space of infrared image and visible images indicates:
L:I(x1,y1)×t1→I(x1,y1;t1), wherein wherein L converts for scale space, x1It indicates in infrared image I
Abscissa, y1Indicate the ordinate in infrared image I, t1Indicate the scale factor of infrared image multiscale space.L:V(x2,
y2)×t2→V(x2,y2;t2), wherein wherein L converts for scale space, x2Indicate the abscissa in visible images V, y2Table
Show the ordinate in visible images V, t2Indicate the scale factor of visible images multiscale space.
For infrared image:I(x1,y1;t1)=I (x1,y1;0)*g(x1,y1;t1), infrared image scaling function
For visible images:V(x2,y2;t2)=V (x2,y2;0)*g(x2,y2;t2), it is seen that light image scaling function
In step 2 of the present invention, the visual attention computation model of infrared image and visible images is established, calculates infrared image
With the vision attention notable figure of visible images, including:Brightness, color and the direction for calculating infrared image and visible images are low
Layer visual signature figure and corresponding each characteristic remarkable picture.
For infrared image:
Brightness notable figure calculation formula:
C indicates that the high yardstick factor, s indicate that low scale factor, I (c) indicate that high yardstick is empty
Between infrared image, I (s) indicates the infrared image of low scale space.
Color notable figure calculation formula:
Red green/green red double opposition images of infrared image
Wherein IR (c) is the high yardstick red channel image of infrared image, and IR (s) is the low scale red channel image of infrared image,
IG (c) is the high yardstick green channel images of infrared image, and IG (s) is the low scale green channel images of infrared image;
The double opposition images of blue Huang/champac of infrared image
Wherein, IB (c) is the high yardstick blue channel image of infrared image, and IB (s) is the low scale blue channel image of infrared image,
IY (c) is the high yardstick yellow channels image of infrared image, and IY (s) is the low scale yellow channels image of infrared image.
Direction notable figure calculation formula:
The direction notable figure of infrared imageWherein, IO (c, θ) is infrared image
High yardstick direction character figure, IO (s, θ) is the low dimension characteristic pattern of infrared image, and θ values are
The vision attention notable figure of entire infrared image is calculated with normalization operator N ():
IS=N (I)+N (IRG)+N (IBY)+N (IO).
For visible images:
Brightness notable figure calculation formula:
C indicates that the high yardstick factor, s indicate that low scale factor, V (c) indicate high yardstick space
Visible images, V (s) indicates the visible images of low scale space.
Color notable figure calculation formula:
Red green/green red double opposition images of visible images
Wherein VR (c) is the high yardstick red channel image of visible images, and VR (s) is the low scale red channel figure of visible images
Picture, VG (c) are the high yardstick green channel images of visible images, and VG (s) is the low scale green channel figure of visible images
Picture;
The double opposition images of blue Huang/champac of visible images
Wherein, VB (c) is the high yardstick blue channel image of visible images, and VB (s) is the low scale blue channel of visible images
Image, VY (c) are the high yardstick yellow channels image of visible images, and VY (s) is the low scale yellow channels of visible images
Image.
Direction notable figure calculation formula:
Wherein, VO (c, θ) is visible to the direction notable figure of visible images
The high yardstick direction character figure of light image, VO (s, θ) are the low dimension characteristic pattern of visible images, and θ values are
The vision attention notable figure of entire visible images is calculated with normalization operator N ():
VS=N (V)+N (VRG)+N (VBY)+N (VO).
In step 3 of the present invention, on the basis of infrared and visible images vision attention notable figure, choose in visual saliency map
The maximum point of gray value is that the gray value of blinkpunkt for the first time is set as by blinkpunkt after selecting first well-marked target region for the first time
Zero, it is second of blinkpunkt to select in visual task notable figure maximum o'clock of gray value again, select successively infrared image and
Well-marked target region in visible images, and M is remembered respectivelyI, MV, MIIndicate the significant target area number in infrared image,
MVIndicate the significant target area in visible images, the well-marked target region M in entire scenesumFor:Msum=MI+MV-
(MI∩MV)。
In step 4 of the present invention, for MsumA salient region, the linear weighted function blending algorithm pair merged using Pixel-level
Well-marked target region carries out fusion treatment, obtains the blending image F of marking areasaliency(i, j)=wI(i,j)·I(i,j)+
wV(i, j) V (i, j)+C, for MIVW is arranged in a region all significant in infrared image and visible imagesI(i, j)=wV
(i, j)=0.5;For MI-IVW is arranged in a region significant only in infrared imageI(i, j)=0.8, wV(i, j)=0.2;It is right
In MV-IVW is arranged in a region significant only in visible imagesI(i, j)=0.2, wV(i, j)=0.8;To non-significant target
Region obtains blending image F using the fusion for carrying out non-significant target area based on principal component analytical methodnon-saliency(i,
j);
In step 5 of the present invention, the infrared image of entire scene and the fusion results of visible images are calculated, F (i, j)=
Fsaliency(i,j)+Fnon-saliency(i,j)。
Advantageous effect:The invention has the advantages that:The multi-level Fusion model of infrared image and visible images is established, is utilized
Visual attention computation model selects the well-marked target in infrared image and visible images respectively, is carried out to well-marked target region
Pixel-level merges, and ensures the fusion mass of infrared image and visible images, and feature is carried out to other non-limiting background areas
Grade fusion, the data volume of processing greatly reduce reason.Method proposed by the present invention while ensuring to well-marked target syncretizing effect,
The fusion efficiencies to a large amount of non-significant target areas are improved, fusion mass and efficiency have been taken into account.
Description of the drawings
The present invention is done with reference to the accompanying drawings and detailed description and is further illustrated, of the invention is above-mentioned
And/or otherwise advantage will become apparent.
Fig. 1 is the infrared and visible light image fusion method flow chart based on well-marked target
Fig. 2 is the infrared image of a certain scene
The visible images that Fig. 3 is with Fig. 2 is Same Scene
Fig. 4 is the blending image that the infrared and visible light image fusion method based on well-marked target obtains
Specific implementation mode
In conjunction with Fig. 1, the present invention is based on the infrared and visible images multi-level Fusion method of well-marked target, specifically include with
Lower step:
Step 1, the multiscale space that infrared image and visible images are established using Nonlinear Scale Space Theory method is indicated.
Infrared image equipped with certain scene is I (x1,y1), (x1,y1∈R2)x1Indicate the abscissa in infrared image I, y1Indicate infrared
Ordinate in image I;Visible images are V (x2,y2), (x2,y2∈R2), x2Indicate the abscissa in visible images V, y2
Indicate the ordinate in visible images V.Scale factor is denoted as t ∈ R+, the scale space table of infrared image and visible images
Show respectively L:I(x1,y1)×t1→I(x1,y1;t1), t1Indicate the scale factor of infrared image multiscale space.L:V(x2,
y2)×t2→V(x2,y2;t2), t2Indicate the scale factor of visible images multiscale space.
I(x1,y1;0)=I (x1,y1)Lt+s(I)=Lt(Ls(I)) (1)
V(x2,y2;0)=V (x2,y2)Lt+s(V)=Lt(Ls(V)) (2)
Wherein L is scale space transformation, and the initial value of scale t is the scale in original image, generally 0.The expansion that L meets
Scattered equation is:
Wherein,Indicate that the equilibrium response of diffusion, D are that a positive definite symmetric matrices indicates diffusion tensor.Such as
FruitWithIt is parallel, then it represents that diffusion is isotropic, otherwise is anisotropy.If D is a normal number, accordingly
Scale space is linear-scale space;If D is one and the relevant scalar function of picture structure, for non-linear isotropism ruler
Spend space;If D be with the relevant vector function of picture structure, for Anisotropic Nonlinear scale space.
The linear-scale space indicated with gaussian kernel function is a kind of Gaussian convolution smoothing process not generating new construction.
For infrared image, can obtain:
I(x1,y1;t1)=I (x1,y1;0)*g(x1,y1;t1) infrared image scaling function
For visible images, can obtain:
V(x2,y2;t2)=V (x2,y2;0)*g(x2,y2;t2) visible images scaling function
Linearity scale space indicate in Gaussian function in smoothed image while noise, some also smooth weights
The feature wanted, to be very difficult to extract these features on thick scale;Meanwhile the use of isotropic linear diffusion equation is led
It is difficult to determine to cause the correspondence at edge between different scale, so that the correspondence between some thickness scales indicate calculates
It is difficult.To make the details such as the object edge in image not be smoothed, the partial zones for the processing image that diffusion equation should be adaptive
Domain, to keep the information for being not intended to be blurred, i.e. method of diffusion should be non-linear.With non-linear isotropic diffusion equation
The Nonlinear Scale Space Theory for establishing image indicates, can ensure that the region contour in each scale image is enhanced, and be conducive to not
It is calculated with the correspondence between scale, shown in diffusion equation such as formula (6):
Or
Wherein, λ>0 indicates edge threshold.
Step 2, it on the basis of image non-linear scale space indicates, is calculated using visual attention computation model infrared and can
The vision attention notable figure of light-exposed image.
On the basis of step 1 indicates infrared image and visible images scale space, color, brightness and side are extracted respectively
To the low-level feature as guiding vision attention, all kinds of characteristic remarkable pictures are calculated by center-periphery difference of receptive field, are finally led to
It crosses normalization operator and obtains vision attention notable figure.
(1) low-level visual feature figure
For infrared image:
Brightness figure:Remember Ir (t1)、Ig(t1) and Ib (t1) it is respectively red, green and indigo plant in original infrared image
Chrominance channel, wherein t1Indicate that scale factor, luminance graph are:
I(t1)=(Ir (t1)+Ig(t1)+Ib(t1))/3 (8)
Color characteristic figure:With I (t1) to r (t1)、g(t1) and b (t1) channel is normalized, can obtain sensu lato
Red, green, blue and Huang Si channel value, such as formula (9), (10), (11), shown in (12).
IR(t1)=Ir (t1)-(Ig(t1)+Ib(t1))/2 (9)
IG(t1)=Ig (t1)-(Ir(t1)+Ib(t1))/2 (10)
IB(t1)=Ib (t1)-(Ir(t1)+Ig(t1))/2 (11)
IY(t1)=Ir (t1)+Ig(t1)-2*(|Ir(t1)-Ig(t1)|+Ib(t1)) (12)
Wherein, negative value is set as 0.
Direction character figure:The direction character figure that infrared image is calculated using Gabor filter, as shown in formula (13):
Wherein (x0,y0) it is receptive field centre coordinate in spatial domain, x0For abscissa, y0For ordinate;(ξ0,υ0) it is filter
Optimal spatial frequency on frequency domain, ξ0For real part, υ0For imaginary part.α and β is the standard of Gaussian function on x and y-axis direction respectively
Difference.The present invention takes the Gabor filter of four direction to export and is used as directional vision characteristic pattern:N is constant 4, and i is indicated not
Same direction, value 0,1,2,3.
For visible images:
Brightness figure:Remember Vr (t2)、Vg(t2) and Vb (t2) be respectively red in primary visible light image, green and
Blue channel, wherein t2Indicate scale factor, then luminance graph is:
V(t2)=(Vr (t2)+Vg(t2)+Vb(t2))/3 (14)
Color characteristic figure:With V (t2) to Vr (t2)、Vg(t2) and Vb (t2) channel is normalized, can obtain broadly
Red, green, blue and Huang Si channel value, such as formula (15), (16), (17), shown in (18).
VR(t2)=Vr (t2)-(Vg(t2)+Vb(t2))/2 (15)
VG(t2)=Vg (t2)-(Vr(t2)+Vb(t2))/2 (16)
VB(t2)=Vb (t2)-(Vr(t2)+Vg(t2))/2 (17)
VY(t2)=Vr (t2)+Vg(t2)-2*(|Vr(t2)-Vg(t2)|+Vb(t2)) (18)
Wherein, negative value is set as 0.
Direction character figure:The direction character figure that visible images are calculated using Gabor filter, as shown in formula (19):
Wherein (x0,y0) it is receptive field centre coordinate in spatial domain, x0For abscissa, y0For ordinate;(ξ0,υ0) it is filter
Optimal spatial frequency on frequency domain, ξ0For real part, υ0For imaginary part.α and β is the standard of Gaussian function on x and y-axis direction respectively
Difference.The present invention takes the Gabor filter of four direction to export and is used as directional vision characteristic pattern:N is constant 4, and i is indicated not
Same direction, value 0,1,2,3.
(2) characteristic remarkable picture
For infrared image:
By brightness I, the method for four color components IR, IG, IB, IY and four direction characteristic pattern scale space carries out table
Show, center image correspond to high-resolution under scale, periphery image correspond to low resolution under scale, realize receptive field and
The calculative strategy for integrating wild center-periphery difference, obtains each characteristic remarkable picture.
Brightness notable figure:
Brightness notable figure is generated by luminance contrast, and brightness notable figure is indicated with I (c, s):
Wherein, c is the scale factor that infrared image scale space indicates middle high-resolution, and s is low resolution scale factor,Son is calculated for center-periphery difference, the reality of the respective pixel in low-resolution image is subtracted by the pixel in high-definition picture
It is existing.
Color notable figure:
Red green/green red double opposition images of infrared image:
Wherein IR (c) is the high yardstick red channel image of infrared image, and IR (s) is that the low scale red of infrared image is logical
Road image, IG (c) are the high yardstick green channel images of infrared image, and IG (s) is the low scale green channel figure of infrared image
Picture.
The double opposition images of blue Huang/champac of infrared image:
Wherein, IB (c) is the high yardstick blue channel image of infrared image, and IB (s) is the low scale blue of infrared image
Channel image, IY (c) are the high yardstick yellow channels image of infrared image, and IY (s) is the low scale yellow channels of infrared image
Image.
Direction notable figure:
Using Gabor pyramids IO (σ, θ), image local directional information can be obtained, wherein σ is scale factor, θ ∈
{0°,45°,90°,135°}.It can be by direction by the calculating of local direction contrast by the calculating of local direction contrast
Notable figure IO (c, s, θ) is encoded into one group:
For visible images:
By brightness V, the method for four color components VR, VG, VB, VY and four direction characteristic pattern scale space carries out table
Show, center image correspond to high-resolution under scale, periphery image correspond to low resolution under scale, realize receptive field and
The calculative strategy for integrating wild center-periphery difference, obtains each characteristic remarkable picture.
Brightness notable figure:
Brightness notable figure is generated by luminance contrast, and brightness notable figure is indicated with V (c, s):
Wherein, c be visible images scale space indicate middle high-resolution scale factor, s be low resolution scale because
Son,Son is calculated for center-periphery difference, the respective pixel in low-resolution image is subtracted by the pixel in high-definition picture
It realizes.
Color notable figure:
Red green/green red double opposition images of visible images:
Wherein VR (c) is the high yardstick red channel image of visible images, and VR (s) is that the low scale of visible images is red
Chrominance channel image, VG (c) are the high yardstick green channel images of visible images, and VG (s) is that the low scale of visible images is green
Chrominance channel image.
The double opposition images of blue Huang/champac of visible images:
Wherein, VB (c) is the high yardstick blue channel image of visible images, and VB (s) is the low scale of visible images
Blue channel image, VY (c) are the high yardstick yellow channels image of visible images, and VY (s) is the low scale of visible images
Yellow channels image.
Direction notable figure:
Using Gabor pyramids VO (σ, θ), image local directional information can be obtained, wherein σ is scale factor, θ ∈
{0°,45°,90°,135°}.It can be by direction by the calculating of local direction contrast by the calculating of local direction contrast
Notable figure VO (c, s, θ) is encoded into one group:
(3) vision attention notable figure
The low-level feature figure of brightness, color and direction has different dynamic range and extraction mechanism, due to all spies
Sign notable figure is combined, only the stronger object of conspicuousness in a small number of figures, may by noise or in big spirogram conspicuousness compared with
Small object is covered.Therefore, the present invention to enhance the less characteristic remarkable picture in notable peak, and is cut using normalization operator N ()
The weak characteristic remarkable picture that there are a large amount of significantly peaks.To each category feature notable figure, the operation of the operator includes:1) spy is normalized
It levies in figure to [0 ..., 1] range, to eliminate the amplitude difference dependent on feature;2) all local poles in addition to global maximum are calculated
Big mean value) useMultiply this feature notable figure.
For infrared image:
Brightness, color and direction character notable figure are normalized respectively with normalization operator N (), obtained WithIt is shown below.
Symbol ⊕ indicates point-by-point summation in formula.
Vision attention notable figure has formula (22) calculating:
For visible images:
Brightness, color and direction character notable figure are normalized respectively with normalization operator N (), obtained WithIt is shown below.
Symbol ⊕ indicates point-by-point summation in formula.
Vision attention notable figure has formula (22) calculating:
Step 3, it on the basis of infrared and visible images vision attention notable figure, is selected respectively using inhibition of return mechanism
Go out the well-marked target region in infrared and visible images, and calculates all well-marked target regions in entire scene.
After visual saliency map is calculated by step 2, it is to watch attentively for the first time to choose the maximum point of gray value in visual saliency map
Point goes out first well-marked target region centered on the point with the rectangular selection of original image size 1/16.First well-marked target region
After selecting, the gray value of blinkpunkt for the first time is set as zero, select again in visual task notable figure the maximum point of gray value for
Second of blinkpunkt goes out second well-marked target region using the rectangular selection of original image size 1/16.Iterative cycles select one by one
Select out the well-marked target region in original image.Well-marked target region in the infrared image and visible images selected is respectively
MI, MV。
Well-marked target region in entire scene is Msum=MI+MV-(MI∩MV)。
Step 4, registration operation is carried out to infrared image and visible images, well-marked target region is merged using Pixel-level
Algorithm carries out fusion treatment, and to non-significant target area, fusion treatment is carried out using feature-based fusion algorithm.
Using crucial point match method, registration operation is carried out to infrared image I and visible images V.
For salient region, it is M to calculate the well-marked target in entire scene by step 3sum, wherein in infrared image
It is M with all significant areal in visible imagesIV=MI∩MV, significant areal is M only in infrared imageI-IV
=MI-MIV, significant areal is M only in visible imagesV-IV=MV-MIV, using the linear weighted function of Pixel-level fusion
Blending algorithm carries out fusion treatment to well-marked target region, obtains the blending image F of marking areasaliency(i, j), with formula (23)
It indicates:
Fsaliency(i, j)=wI(i,j)·I(i,j)+wV(i,j)·V(i,j)+C (23)
Wherein, I (i, j) and V (i, j) indicates that the pixel position in source images, F (i, j) indicate the picture in blending image
Vegetarian refreshments position.wI(i, j) and wV(i, j) is weights, and wI(i,j)+wV(i, j)=1.For MIVIt is a in infrared image and visible
W is arranged in all significant region in light imageI(i, j)=wV(i, j)=0.5;For MI-IVIt is a significant only in infrared image
W is arranged in regionI(i, j)=0.8, wV(i, j)=0.2;For MV-IVW is arranged in a region significant only in visible imagesI
(i, j)=0.2, wV(i, j)=0.8.
To non-significant target area, feature level figure is used to the non-significant target area in infrared image and visible images
As blending algorithm carries out fusion treatment.Using based on principal component analysis (Principal Component in this patent
Analysis, PCA) method carries out the fusion of non-significant target area.
Infrared image and visible images are converted by PCA shift theories respectively, obtain each principal component, by infrared figure
The first principal component image of picture and visible images carries out Histogram Matching, then replaces the first Main classification with full-colour image,
Blending image F is being obtained by PCA inverse transformationsnon-saliency(i,j)。
Step 5, comprehensive well-marked target region fusion results and non-significant region fusion results, generate the infrared figure of whole scene
Picture and visual image fusion image.
The blending image F in well-marked target region in step 4 calculates scenesaliency(i, j) and non-significant target area
Image Fnon-saliencyOn the basis of (i, j), the infrared image of entire scene and the fusion results of visible images are calculated, such as formula
(24) shown in.
F (i, j)=Fsaliency(i,j)+Fnon-saliency(i,j) (24)
Embodiment
Illustrate the realization process of the present invention by a specific example.
Fig. 2 is the infrared image of a certain scene, and Fig. 3 is the visible images of a certain scene.
It is indicated according to the Nonlinear Scale Space Theory for described in step 1, initially setting up infrared image and visible images, edge threshold
Value λ is taken as 0.5.
According to brightness, color and the directional vision feature for described in step 2, calculating separately infrared image and visible images
Figure and brightness, color and direction notable figure calculate the vision attention notable figure of infrared image and visible images.
It it is 5 according to the well-marked target region described in step 3, calculated in infrared image, it is seen that notable in light image
Target area is 4, wherein all significant target is 3 in infrared image and visible images, the notable mesh in entire scene
It is 6 to mark region.
According to described in step 4,6 well-marked target regions in scene are merged according to pixel level fusing method, it is right
It is merged according to feature level fusing method other non-significant target areas.
According to described in step 5, the fusion results in well-marked target region and the fusion results of non-significant target area are carried out
It is comprehensive, obtain entire scene infrared image and visual image fusion as a result, as shown in Figure 4.
Effect of the present invention can be summarised as:
1) the multi-level Fusion model for establishing image co-registration, in scene well-marked target and non-significant target be respectively adopted
Pixel level fusing method and feature level fusing method carry out Pixel-level fusion to well-marked target, i.e. essence fusion, to other background areas
Domain carries out feature-based fusion, i.e., thick fusion improves while ensuring to well-marked target syncretizing effect to a large amount of non-significant targets
The fusion efficiencies in region meet with the visual cognition process height of the mankind.
2) the multiscale space expression of image, the noise in smoothed image are established using Nonlinear Scale Space Theory method
Simultaneously, it is ensured that well-marked target profile is enhanced in each scale image, improves the fusion mass to well-marked target in scene.
3) visual attention computation model for establishing infrared image and visible images, in terms of brightness, color and direction three
Metric objective conspicuousness, and calculated separately in infrared image and visible images by normalizing operator and inhibition of return mechanism
Well-marked target, simultaneous selection goes out the well-marked target in infrared image and visible images, and carries out fusion treatment.
4) method proposed by the present invention cannot be only used for the fusion treatment of infrared image and visible images, can also be into one
Step is generalized to other multi-source image fusions field, such as multispectral image and visual image fusion, SAR image and visible light figure
Picture, SAR image and infrared image etc., application prospect is extensive.
The present invention provides a kind of infrared and visible light image fusion method based on well-marked target, implements the technology
There are many method and approach of scheme, the above is only a preferred embodiment of the present invention, it is noted that for the art
Those of ordinary skill for, various improvements and modifications may be made without departing from the principle of the present invention, these change
Protection scope of the present invention is also should be regarded as into retouching.The available prior art of each component part being not known in the present embodiment adds
To realize.
Claims (1)
1. a kind of infrared and visible light image fusion method based on well-marked target, which is characterized in that include the following steps:
Step 1, the non-linear multiscale space of infrared image and visible images is established using non-linear multiscale space method
It indicates:
If being I (x there are one the infrared image of scene1,y1), x1Indicate the abscissa in infrared image I, y1Indicate infrared image I
In ordinate;Visible images are V (x2,y2), x2Indicate the abscissa in visible images V, y2Indicate visible images V
In ordinate, scale factor is denoted as t ∈ R+, the non-linear multiscale space expression of infrared image and visible images is respectively
I(x1,y1)×t1→I(x1,y1;t1), t1Indicate the scale factor of the non-linear multiscale space of infrared image, V (x2,y2)×t2
→V(x2,y2;t2), t2Indicate the scale factor of the non-linear multiscale space of visible images:
I(x1,y1;0)=I (x1,y1)
V(x2,y2;0)=V (x2,y2)
Wherein I (x1,y1;0) it is the infrared image of original scale, V (x2,y2;0) it is the visible images of original scale, scale t's
Initial value is the scale in original image;
The linear-scale space indicated with gaussian kernel function is the Gaussian convolution smoothing process for not generating new construction;
For infrared image, obtain:
I(x1,y1;t1)=I (x1,y1;0)*g(x1,y1;t1)
Infrared image scaling function
For visible images, obtain:
V(x2,y2;t2)=V (x2,y2;0)*g(x2,y2;t2)
Visible images scaling function
The regional area of the adaptive processing image of diffusion equation, to keep the information for being not intended to be blurred, i.e. method of diffusion
It is non-linear;It is indicated with the non-linear multiscale space of non-linear isotropic diffusion establishing equation image, it is ensured that each scale
Region contour in image is enhanced, and is conducive to the calculating of the correspondence between different scale, and diffusion equation is shown below:
Or
Wherein,It is expressed as the gradient of image,Indicate the modulus value of the gradient of image,The diffusion coefficient of expression, λ
> 0 indicates edge threshold;
Step 2, it on the basis of image non-linear multiscale space indicates, is calculated using visual attention computation model infrared and visible
The vision attention notable figure of light image;
On the basis of step 1 indicates infrared image and visible images scale space, color, brightness and direction are extracted respectively and is made
To guide the low-level feature of vision attention, all kinds of characteristic remarkable pictures are calculated by the center of receptive field-periphery difference, finally by returning
One change operator obtains vision attention notable figure;
(1) low-level visual feature figure:
For infrared image:
Brightness figure:Remember Ir (t1)、Ig(t1) and Ib (t1) it is respectively that red in original infrared image, green and blue are logical
Road, wherein t1Indicate that scale factor, brightness figure are:
I(t1)=(Ir (t1)+Ig(t1)+Ib(t1))/3
Color characteristic figure:To I (t1) r (t1)、g(t1) and b (t1) channel is normalized, obtain sensu lato red IR (t1)、
Green IG (t1), indigo plant IB (t1) and Huang IY (t1) four channel values, as shown in following formula:
IR(t1)=Ir (t1)-(Ig(t1)+Ib(t1))/2
IG(t1)=Ig (t1)-(Ir(t1)+Ib(t1))/2
IB(t1)=Ib (t1)-(Ir(t1)+Ig(t1))/2
IY(t1)=Ir (t1)+Ig(t1)-2*(|Ir(t1)-Ig(t1)|+Ib(t1))
Wherein, it is set as 0 if channel value result is negative value;
Direction character figure:The direction character figure that infrared image is calculated using Gabor filter, is shown below:
Wherein (x1,y1) be infrared image pixel coordinate, (x0,y0) it is receptive field centre coordinate in spatial domain, x0For abscissa, y0
For ordinate;(ξ0,υ0) it is optimal spatial frequency of the filter on frequency domain, ξ0For real part, υ0For imaginary part, i1 indicates different
Direction;σ and β is the standard deviation of Gaussian function on x and y-axis direction respectively;The Gabor filter of four direction is taken to export conduct side
To characteristic pattern:Four direction is respectively
For visible images:
Brightness figure:Remember Vr (t2)、Vg(t2) and Vb (t2) it is respectively red, green and blue in primary visible light image
Channel, wherein t2Indicate scale factor, then brightness figure is:
V(t2)=(Vr (t2)+Vg(t2)+Vb(t2))/3
Color characteristic figure:To V (t2) Vr (t2)、Vg(t2) and Vb (t2) channel is normalized, obtain it is sensu lato it is red, green,
Blue and four channel value VR (t of Huang2), VR (t2), VB (t2), VY (t2), as shown in following formula:
VR(t2)=Vr (t2)-(Vg(t2)+Vb(t2))/2
VG(t2)=Vg (t2)-(Vr(t2)+Vb(t2))/2
VB(t2)=Vb (t2)-(Vr(t2)+Vg(t2))/2
VY(t2)=Vr (t2)+Vg(t2)-2*(|Vr(t2)-Vg(t2)|+Vb(t2))
Wherein, it is set as 0 if channel value result is negative value;
Direction character figure:The direction character figure that visible images are calculated using Gabor filter, is shown below:
Wherein (x2,y2) be visible images pixel coordinate;(x0,y0) it is receptive field centre coordinate in spatial domain, x0For abscissa,
y0For ordinate;(ξ0,υ0) it is optimal spatial frequency of the filter on frequency domain, ξ0For real part, υ0For imaginary part, i1 indicates different
Direction;σ and β is the standard deviation of Gaussian function on x and y-axis direction respectively;The Gabor filter of four direction is taken to export conduct side
To characteristic pattern:Four direction is respectively
(2) characteristic remarkable picture
For infrared image:
By brightness I, the method for four color components IR, IG, IB, IY and four direction characteristic pattern scale space is indicated,
Center image corresponds to the scale under high-resolution, and periphery image corresponds to the scale under low resolution, realizes receptive field and whole
The calculative strategy for closing wild center-periphery difference, obtains each characteristic remarkable picture;
Brightness notable figure:
Brightness notable figure is generated by luminance contrast, and brightness notable figure is indicated with I (c, s):
Wherein, c is the scale factor that infrared image scale space indicates middle high-resolution, and s is low resolution scale factor,For
Center-periphery difference calculates son, and subtracting the respective pixel in low-resolution image by the pixel in high-definition picture realizes;
Color notable figure:
Red green/green red double opposition images of infrared image:
Wherein IR (c) is the high yardstick red channel image of infrared image, and IR (s) is the low scale red channel figure of infrared image
Picture, IG (c) are the high yardstick green channel images of infrared image, and IG (s) is the low scale green channel images of infrared image;
The double opposition images of blue Huang/champac of infrared image:
Wherein, IB (c) is the high yardstick blue channel image of infrared image, and IB (s) is the low scale blue channel of infrared image
Image, IY (c) are the high yardstick yellow channels image of infrared image, and IY (s) is the low scale yellow channels image of infrared image;
Direction notable figure:
Image local directional information is obtained using Gabor pyramids IO (σ 1, θ), wherein σ 1 is scale factor, θ ∈ 0 °, 45 °,
90°,135°};By the calculating of local direction contrast, direction notable figure IO (c, s, θ) is encoded into one group:
IO (s, θ) | indicate that the direction character figure of low-resolution image, IO (c, θ) indicate the direction character figure of high-definition picture;
For visible images:
By brightness V, the method for four color components VR, VG, VB, VY and four direction characteristic pattern scale space is indicated,
Center image corresponds to the scale under high-resolution, and periphery image corresponds to the scale under low resolution, realizes receptive field and whole
The calculative strategy for closing wild center-periphery difference, obtains each characteristic remarkable picture;
Brightness notable figure:
Brightness notable figure is generated by luminance contrast, and brightness notable figure is indicated with V (c, s):
Wherein, c is the scale factor that visible images scale space indicates middle high-resolution, and s is low resolution scale factor,
Son is calculated for center-periphery difference, the reality of the respective pixel in low-resolution image is subtracted by the pixel in high-definition picture
It is existing;
Color notable figure:
Red green/green red double opposition images of visible images:
Wherein VR (c) is the high yardstick red channel image of visible images, and VR (s) is that the low scale red of visible images is logical
Road image, VG (c) are the high yardstick green channel images of visible images, and VG (s) is that the low scale green of visible images is logical
Road image;
The double opposition images of blue Huang/champac of visible images:
Wherein, VB (c) is the high yardstick blue channel image of visible images, and VB (s) is the low scale blue of visible images
Channel image, VY (c) are the high yardstick yellow channels image of visible images, and VY (s) is the low scale yellow of visible images
Channel image;
Direction notable figure:
Using Gabor pyramids VO (σ 1, θ), image local directional information, wherein σ are obtained1For scale factor, θ ∈ 0 °, 45 °,
90°,135°};By the calculating of local direction contrast, direction notable figure VO (c, s, θ) can be encoded into one group:
(3) vision attention notable figure:
Using normalization operator N (), to enhance the less characteristic remarkable picture in notable peak, and weaken the spy that there are a large amount of significantly peaks
Notable figure is levied, to each category feature notable figure, the operation of the operator includes:
1) it normalizes in this feature notable figure to 0~1 range, to eliminate the amplitude difference dependent on feature;
2) mean value of all local maximums in addition to global maximum is calculated
3) it usesMultiply this feature notable figure;
For infrared image:
Brightness, color and direction notable figure are normalized respectively with normalization operator N (), obtained WithIt is shown below:
Symbol in formulaIndicate point-by-point summation;
Infrared image vision attention notable figure is calculated by the following formula:
For visible images:
Brightness, color and direction notable figure are normalized respectively with normalization operator N (), obtained
WithIt is shown below:
Symbol in formulaIndicate point-by-point summation;
Visible images vision attention notable figure is calculated by the following formula:
Step 3, it on the basis of infrared and visible images vision attention notable figure, is selected respectively using inhibition of return mechanism red
Well-marked target region in outer and visible images, and calculate all well-marked target regions in entire scene;
After vision attention notable figure is calculated by step 2, it is for the first time to choose the maximum point of gray value in vision attention notable figure
Blinkpunkt goes out first well-marked target region, first well-marked target centered on the point with the rectangular selection of original image size 1/16
After regional choice goes out, the gray value of blinkpunkt for the first time is set as zero, selects gray value in visual task notable figure maximum again
Point is second of blinkpunkt, goes out second well-marked target region using the rectangular selection of original image size 1/16, iterative cycles, by
A well-marked target region selected in original image, the well-marked target region point in the infrared image and visible images selected
It Wei not MI, MV;
Well-marked target region in entire scene is Msum=MI+MV-(MI∩MV);
Step 4, registration operation is carried out to infrared image and visible images, Pixel-level blending algorithm is used to well-marked target region
Fusion treatment is carried out, to non-significant target area, fusion treatment is carried out using feature-based fusion algorithm;
Using crucial point match method, registration operation is carried out to infrared image I and visible images V;
For well-marked target region, it is M to calculate the well-marked target region in entire scene by step 3sum, wherein in infrared figure
All significant areal is M in picture and visible imagesIV=MI∩MV, significant areal is only in infrared image
MI-IV=MI-MIV, significant areal is M only in visible imagesV-IV=MV-MIV, using Pixel-level fusion it is linear plus
It weighs blending algorithm and fusion treatment is carried out to well-marked target region, obtain the blending image F in well-marked target regionsaliency(i, j) is used
Following formula indicates:
Fsaliency(i, j)=wI(i,j)·I(i,j)+wV(i,j)·V(i,j)+C
Wherein, I (i, j) and V (i, j) indicates that the pixel position in source images, F (i, j) indicate the pixel in blending image
Position;wI(i, j) and wV(i, j) is weights, and wI(i,j)+wV(i, j)=1;For MIVIt is a in infrared image and visible light figure
W is arranged in all significant region as inI(i, j)=wV(i, j)=0.5;For MI-IVA region significant only in infrared image,
W is setI(i, j)=0.8, wV(i, j)=0.2;For MV-IVW is arranged in a region significant only in visible imagesI(i,j)
=0.2, wV(i, j)=0.8;
To non-significant target area, the non-significant target area in infrared image and visible images is melted using feature level image
Hop algorithm carries out fusion treatment, using the fusion for carrying out non-significant target area based on principal component analytical method;
Infrared image and visible images are converted by PCA shift theories respectively, obtain each principal component, by infrared image and
The first principal component image of visible images carries out Histogram Matching, then replaces first principal component with full-colour image, then lead to
It crosses PCA inverse transformations and obtains blending image Fnon-saliency(i,j);
Step 5, comprehensive well-marked target region fusion results and non-significant target area fusion results, generate the infrared figure of whole scene
Picture and visual image fusion image;
The blending image F in well-marked target region in step 4 calculates scenesaliency(i, j) and non-significant target area image
Fnon-saliencyOn the basis of (i, j), the infrared image of entire scene and the fusion results of visible images are calculated, such as following formula institute
Show:
F (i, j)=Fsaliency(i,j)+Fnon-saliency(i,j)。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510111415.0A CN104700381B (en) | 2015-03-13 | 2015-03-13 | A kind of infrared and visible light image fusion method based on well-marked target |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510111415.0A CN104700381B (en) | 2015-03-13 | 2015-03-13 | A kind of infrared and visible light image fusion method based on well-marked target |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104700381A CN104700381A (en) | 2015-06-10 |
CN104700381B true CN104700381B (en) | 2018-10-12 |
Family
ID=53347469
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510111415.0A Active CN104700381B (en) | 2015-03-13 | 2015-03-13 | A kind of infrared and visible light image fusion method based on well-marked target |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104700381B (en) |
Families Citing this family (22)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11010876B2 (en) * | 2015-06-26 | 2021-05-18 | Nec Corporation | Image processing system, image processing method, and computer-readable recording medium |
CN106251355B (en) * | 2016-08-03 | 2018-12-14 | 江苏大学 | A kind of detection method merging visible images and corresponding night vision infrared image |
CN106530266B (en) * | 2016-11-11 | 2019-11-01 | 华东理工大学 | A kind of infrared and visible light image fusion method based on region rarefaction representation |
CN106898008A (en) * | 2017-03-01 | 2017-06-27 | 南京航空航天大学 | Rock detection method and device |
CN107292872A (en) * | 2017-06-16 | 2017-10-24 | 艾松涛 | Image processing method/system, computer-readable recording medium and electronic equipment |
CN107918748A (en) * | 2017-10-27 | 2018-04-17 | 南京理工大学 | A kind of multispectral two-dimension code recognition device and method |
CN108198157A (en) * | 2017-12-22 | 2018-06-22 | 湖南源信光电科技股份有限公司 | Heterologous image interfusion method based on well-marked target extracted region and NSST |
CN108288344A (en) * | 2017-12-26 | 2018-07-17 | 李文清 | A kind of efficient forest fire early-warning system |
CN108090888B (en) * | 2018-01-04 | 2020-11-13 | 北京环境特性研究所 | Fusion detection method of infrared image and visible light image based on visual attention model |
CN108769550B (en) * | 2018-05-16 | 2020-07-07 | 中国人民解放军军事科学院军事医学研究院 | Image significance analysis system and method based on DSP |
CN109255793B (en) * | 2018-09-26 | 2019-07-05 | 国网安徽省电力有限公司铜陵市义安区供电公司 | A kind of monitoring early-warning system of view-based access control model feature |
CN109447909A (en) * | 2018-09-30 | 2019-03-08 | 安徽四创电子股份有限公司 | The infrared and visible light image fusion method and system of view-based access control model conspicuousness |
CN109493309A (en) * | 2018-11-20 | 2019-03-19 | 北京航空航天大学 | A kind of infrared and visible images variation fusion method keeping conspicuousness information |
CN110210407A (en) * | 2019-06-04 | 2019-09-06 | 武汉科技大学 | A kind of Misty Image well-marked target detection method |
CN110489792B (en) * | 2019-07-12 | 2023-03-31 | 中国人民解放军92942部队 | Method and device for designing visible light camouflage distance and server |
CN111161356B (en) | 2019-12-17 | 2022-02-15 | 大连理工大学 | Infrared and visible light fusion method based on double-layer optimization |
CN111080724B (en) * | 2019-12-17 | 2023-04-28 | 大连理工大学 | Fusion method of infrared light and visible light |
CN110874827B (en) * | 2020-01-19 | 2020-06-30 | 长沙超创电子科技有限公司 | Turbulent image restoration method and device, terminal equipment and computer readable medium |
CN111914422B (en) * | 2020-08-05 | 2021-02-02 | 北京开云互动科技有限公司 | Real-time visual simulation method for infrared features in virtual reality |
CN113159229B (en) * | 2021-05-19 | 2023-11-07 | 深圳大学 | Image fusion method, electronic equipment and related products |
US11967102B2 (en) | 2021-07-16 | 2024-04-23 | Shanghai United Imaging Intelligence Co., Ltd. | Key points detection using multiple image modalities |
CN116977154B (en) * | 2023-09-22 | 2024-03-19 | 南方电网数字电网研究院有限公司 | Visible light image and infrared image fusion storage method, device, equipment and medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101873440A (en) * | 2010-05-14 | 2010-10-27 | 西安电子科技大学 | Infrared and visible light video image fusion method based on Surfacelet conversion |
CN103366353A (en) * | 2013-05-08 | 2013-10-23 | 北京大学深圳研究生院 | Infrared image and visible-light image fusion method based on saliency region segmentation |
CN104408700A (en) * | 2014-11-21 | 2015-03-11 | 南京理工大学 | Morphology and PCA (principal component analysis) based contourlet fusion method for infrared and visible light images |
-
2015
- 2015-03-13 CN CN201510111415.0A patent/CN104700381B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101873440A (en) * | 2010-05-14 | 2010-10-27 | 西安电子科技大学 | Infrared and visible light video image fusion method based on Surfacelet conversion |
CN103366353A (en) * | 2013-05-08 | 2013-10-23 | 北京大学深圳研究生院 | Infrared image and visible-light image fusion method based on saliency region segmentation |
CN104408700A (en) * | 2014-11-21 | 2015-03-11 | 南京理工大学 | Morphology and PCA (principal component analysis) based contourlet fusion method for infrared and visible light images |
Non-Patent Citations (3)
Title |
---|
Visual attention guided image fusion with sparse representation;Bin Yang et al.;《Optik》;20141231;第125卷(第17期);全文 * |
协同视觉选择注意计算模型研究;邵静;《中国博士学位论文全文数据库 信息科技辑》;20081115;第2008年卷(第11期);第51-52、55-57、62页 * |
基于仿生视觉机理的多源图像融合;万莉;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140815;第2014年卷(第8期);第36、41、47、48、52页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104700381A (en) | 2015-06-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104700381B (en) | A kind of infrared and visible light image fusion method based on well-marked target | |
Hwang et al. | Multispectral pedestrian detection: Benchmark dataset and baseline | |
Cai et al. | Dehazenet: An end-to-end system for single image haze removal | |
CN109344701A (en) | A kind of dynamic gesture identification method based on Kinect | |
Luo et al. | Multi-scale traffic vehicle detection based on faster R–CNN with NAS optimization and feature enrichment | |
Lalonde et al. | Estimating the natural illumination conditions from a single outdoor image | |
Lu et al. | Multi-scale strip pooling feature aggregation network for cloud and cloud shadow segmentation | |
Luo et al. | Thermal infrared image colorization for nighttime driving scenes with top-down guided attention | |
Renza et al. | A new approach to change detection in multispectral images by means of ERGAS index | |
CN102314602B (en) | Shadow removal in image captured by vehicle-based camera using optimized oriented linear axis | |
CN102314600A (en) | Be used for the shade of the removal of clear path detection by the image of catching based on the camera of vehicle | |
CN104504744B (en) | A kind of true method and device of plan for synthesizing license plate image | |
CN105678318B (en) | The matching process and device of traffic sign | |
CN109816694A (en) | Method for tracking target, device and electronic equipment | |
CN116091372B (en) | Infrared and visible light image fusion method based on layer separation and heavy parameters | |
CN109753945A (en) | Target subject recognition methods, device, storage medium and electronic equipment | |
Sun et al. | IRDCLNet: Instance segmentation of ship images based on interference reduction and dynamic contour learning in foggy scenes | |
Jin et al. | Vehicle license plate recognition for fog‐haze environments | |
Jain et al. | Multi-sensor image fusion using intensity hue saturation technique | |
Xu et al. | COCO-Net: A dual-supervised network with unified ROI-loss for low-resolution ship detection from optical satellite image sequences | |
Maxwell et al. | Real-time physics-based removal of shadows and shading from road surfaces | |
Ying et al. | Region-aware RGB and near-infrared image fusion | |
Shibata et al. | Unified image fusion framework with learning-based application-adaptive importance measure | |
Zhao et al. | Infrared and visible imagery fusion based on region saliency detection for 24-hour-surveillance systems | |
Li et al. | Cloud detection from remote sensing images by cascaded U-shape attention networks |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |