CN101056353A - Infrared focal plane asymmetric correction method based on the motion detection guidance - Google Patents
Infrared focal plane asymmetric correction method based on the motion detection guidance Download PDFInfo
- Publication number
- CN101056353A CN101056353A CN 200710051920 CN200710051920A CN101056353A CN 101056353 A CN101056353 A CN 101056353A CN 200710051920 CN200710051920 CN 200710051920 CN 200710051920 A CN200710051920 A CN 200710051920A CN 101056353 A CN101056353 A CN 101056353A
- Authority
- CN
- China
- Prior art keywords
- picture dot
- image
- bad
- max
- focal plane
- 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.)
- Granted
Links
Images
Landscapes
- Closed-Circuit Television Systems (AREA)
Abstract
An infrared focal plane nonuniformity correction method based on movement detection guidance belongs to the field of infrared focal plane detector and has an objective that it is self-adaptive, can effectively inhibit object degeneration and resolve problems of artifacts (ghosting artifact), and is easy to be realized through hardware. The present invention comprises (1) a pretreatment step, (2) a correction step, (3) an iterative step length adjusting step and (4) a plus correction cofficient renewal step. The present invention effectively resolves and depresses object degeneration and artifacts phenomena generated in scenes arithmetic, eliminates fixing pattern noises, and exhibits adaptability, wherein the calculation operations are simple and easy to be realized through hardware, achieving real-time performances.
Description
Technical field
The invention belongs to the infrared focal plane detector field, be specifically related to a kind of infrared focal plane asymmetric correction method based on motion detection guidance.
Background technology
Infrared focal plane array IRFPA (Infrared Focal Plane Array) imaging system is the direction of infrared imagery technique development, is the core devices of modern infrared imaging system.In first generation infrared imaging system, adopt detector array, be embodied as picture by the one dimension optical mechaical scanning.Compare with first generation infrared imaging system, infrared focal plane array is worked in the mode of staring, and the optical element that does not need to move carries out one dimension or two-dimentional optical mechaical scanning to scenery.Therefore have simple in structure, working stability, spatial resolution height, strong, the frame frequency advantages of higher of detectivity.Staring infrared imaging system has begun to be widely used in civil areas such as night vision, sea rescue search, astronomy, industrial hot-probing and medical science at present, is the developing direction of infrared imaging system.Yet since aspects such as manufactured materials, technology and operational environment, infrared focal plane array ubiquity heterogeneity problem.Its heterogeneity and invalid picture dot have had a strong impact on the image quality of system, have reduced the right metric of spatial resolution, temperature resolution, detection range and the amount of radiation of system, are directly restricting the final performance of system.Although along with the improvement of device making technics, the heterogeneity of focal plane and invalid picture dot have had bigger improvement, from the in addition very big distance of dealing with problems fully, are still the matter of utmost importance that current infrared focal plane array image-forming system must solve.
The asymmetric correction method of present existing infrared focal plane array, mainly be divided into two big classes: a class is based on the bearing calibration of calibration, and this class methods principle is succinct, is easy to hardware realization and integrated; The correction accuracy height can be used for the tolerance of scene temperature, and target is had no requirement, and is the main method that adopts in the actual IRFPA subassembly product.But these class methods are subject to the correction error that the IRFPA response drift is brought; Actual timing needs reference source to demarcate, and makes the apparatus relative complex; Simultaneously need in actual applications periodically to calibrate, beacon frequency depends on the stability of system, is difficult for accomplishing fast reaction for actual detector.The another kind of self-adapting correction method that is based on the scene class, as time domain high-pass filtering correction method, neural net correction method and constant statistical restraint correction method etc.These class methods can overcome the correction error that the IRFPA response drift is brought to a certain extent, do not require or only need simple calibration, according to the adaptive renewal correction coefficient of scene information, become the important research direction of present algorithm research and system applies.
But be based on the scene class methods and all have pseudomorphism (ghosting artifact, be also referred to as " ghost ", promptly be the ghost image of the same scene with it of appearance around real scene) problem, as traditional neural net correcting algorithm, its method combines the characteristics that (is desired output with neighbours' domain space average) and time domain processing (iteration that realizes correction coefficient with the neural net Error Feedback is upgraded) handled in the spatial domain, realize Nonuniformity Correction adaptively, but it requires target to be in continuous motion state, in case it is static that target is tending towards, then target fade-out (fade-out can appear, show that target image fogs, the signal to noise ratio that promptly is target reduces, diminish with the contrast of background, be melted into background, be difficult for being identified); And after target is left owing to motion, can stay the next one to be the pseudomorphism (ghost) of reverted image again in situ.
Narendra is at " Shutterless fixed pattern noise correction for infraredimaging arrays " (Proc.SPIE, 1981,282:44-51) during statistic asymmetric correction method such as proposition, once a kind of simple method of principle ground proposition overcame " ghost ", promptly stopped the renewal of correction coefficient when scenery is constant.
People such as Harris were at " Minimizing the ghosting artifact in scene-basednonuniformity correction " (Proc.SPIE in 1997,1998,3377:106-113) method that in the literary composition above-mentioned Narendra is proposed is realized, be provided with a threshold value in the renewal link, when the variation of certain pixel surpasses this threshold value, can upgrade its correction coefficient, then stop when being lower than this threshold value upgrading.
Liu in 2003 people such as can be led to and analyzed the defective that above-mentioned two pieces of articles are handled, in " analysis of " ghost " problem in the adaptive nonuniformity correction " (infrared technique, 2003,25 (5): 30-32)-proposed to get some weak point of the mode of threshold value among the Wen based on motion, mainly contain following two aspects: (1) threshold value is difficult for determining; (2) spatial noise is constant substantially, can not proofread and correct---and can be understood as the bad correction of fixed pattern noise here, because fixed pattern noise is constant " diagram noise " basically, it moves hardly.
People such as Esteban are at " Ghosting reduction in adaptive nonuniformitycorrection of infrared focal-plane array image sequences " (IEEE, Imageprocessing, 2003, vol.2:II-1001-4) propose in the literary composition to carry out " ghosting reduction " by the method that changes iteration step length, iteration coefficient and 3 * 3 neighborhoods are interior is varied to inverse ratio.That is to say that changing fast local iteration in the neighborhood gets slowly; And neighborhood changes slow local iteration quickening, and wherein the speed of neighborhood variation is decided by the variance of neighborhood gray value.This method has fundamentally also just been kept marginal information, prevents target fade-out, inreal pseudomorphism (ghost) problem that solves.
Zhao in 2004 also people such as worker also in " based on the new algorithm of the infrared focus plane nonuniformity correction of neural net " (infrared technique, 2004,26 (2): 44-47) mention " if scene is static for a long time, then great majority all will lose efficacy based on the method for scene " in the literary composition.Their way be at first with scene first frame image when static preserve, judge then whether scene moves, if not motion just replaces with the image of preserving.If it is static further to be improved to scene then, just calculate the correction coefficient of neural net method as " desired value " with the first frame still image of preserving.
Make a general survey of the present Research of pseudomorphism (ghost) problem of eliminating over the years in the scene class correcting algorithm, most of researchers start with from the motion conditions of scene, but still exist some problems and difficult point not to solve:
(1) step of motion detection can not be too complicated, otherwise be difficult to reach real-time, and making should be consuming time too much for complementary pretreated Nonuniformity Correction work, and a presumptuous guest usurps the role of the host.
(2) judgment threshold of motion detection is difficult for determining.
(3) based on the mode of motion detection class, cause spatial noise constant substantially, can not proofread and correct, because fixed pattern noise is constant " diagram noise " basically, it moves hardly.
Summary of the invention
The invention provides a kind of infrared focal plane asymmetric correction method based on motion detection guidance, purpose is can self adaptation, effectively suppresses target fade-out, solves pseudomorphism (ghost) problem, and be easy to hardware and realize, the heterogeneity of infrared focus plane is proofreaied and correct.
A kind of infrared focal plane asymmetric correction method of the present invention based on motion detection guidance, order comprises:
(1) pre-treatment step, the image with the even irradiation of infrared focal plane detector collection M frame carries out time domain average to them then and obtains the background frames image B, its each picture dot point gray value B
I, jUtilize the invalid picture dot detection algorithm based on scene to carry out invalid picture dot detection to each picture dot point of background frames image, testing result deposits bad meta template in, calculates the average gray B of effective picture dot in the background frames image according to testing result
AveSetting original sequence totalframes to be corrected is N, and N is a natural number, the gain correction coefficient of initialization each picture dot of original image to be corrected
Initialization original sequence n=1 to be corrected, M are 10~100;
(2) aligning step is imported n frame original image V
n, its each picture dot point gray value is V
I, j n, with V
I, j nThe gray value B of subtracting background frame correspondence position
I, j, obtain n frame image X to be corrected
n, its each picture dot point gray value is X
I, j n:
Subscript i wherein, j is respectively abscissa and the ordinate of picture dot in image;
Read bad meta template, treat correcting image X
nThe picture dot of effective picture dot position is proofreaied and correct in the corresponding bad meta template, and updating formula is:
Wherein, G
I, j nIt is n frame gain correction coefficient; Treat correcting image X simultaneously
nThe pixel value of corresponding invalid picture dot position substitutes with the mean value of left neighborhood picture dot pixel and last field picture dot pixel sum, obtains proofreading and correct back image Y at last
n, its each picture dot gray value is Y
I, j n
(3) iteration step length set-up procedure judges whether n>1, otherwise changes step (4); Be that two frames are proofreaied and correct the corresponding picture dot motion of back image variances sigma before and after then calculating
T, ij 2Judge whether
Otherwise adjust iteration step length μ
I, j n, change step (4), be then initialization once more
Original sequence n to be corrected adds 1, judges whether n>N, is then to finish, otherwise changes step (2), wherein motion variance threshold values σ
Tmax 2Be (0.5~1) B
Ave
(4) gain correction coefficient step of updating utilizes the neural net correcting algorithm that the gain heterogeneity of image is compensated, and promptly upgrades gain correction coefficient, and formula is:
F wherein
I, j nFor proofreading and correct back image Y
nIn the neighbours territory picture dot gray value average of each picture dot, after gain correction coefficient upgraded, original sequence n to be corrected added 1, judges whether n>N, is then to finish, otherwise changeed step (2).
Described infrared focal plane asymmetric correction method based on motion detection guidance is characterized in that, in the described pre-treatment step, invalid picture dot testing process is:
(1) calculate the one dimension operator of background frames image in X, Y direction respectively, computing formula is:
bad
xmax=max{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1},
bad
xmin=min{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1},
bad
ymax=max{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j},
bad
ymin=min{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j};
(2) each picture dot point of background frames image is judged:
(B
i,j-bad
xmax)/B
i,j>badTH、(B
i,j-bad
ymax)/B
i,j>badTH、
(bad
Xmin-B
I, j)/B
I, j>badTH or (bad
Ymin-B
I, j)/B
I, j>badTH,
This picture dot of background frames image B B is then judged in arbitrary establishment in four formulas
I, jBe invalid picture dot, otherwise be effective picture dot.
Described infrared focal plane asymmetric correction method based on motion detection guidance is characterized in that, in the described iteration step length set-up procedure:
(1) two frames are proofreaied and correct the corresponding picture dot motion of back image variances sigma before and after
T, ij 2For:
Y wherein
I, j N-1Be that the n-1 frame is proofreaied and correct back image coordinate i, the gray value of j picture dot,
Be that two frames are proofreaied and correct back image coordinate i, the gray average of j picture dot;
(2) adjust iteration step length μ
I, j nProcess is:
Make interim iteration step length
If μ
I, j'>μ
Max, then
If μ
I, j'≤μ
Max, then
μ wherein
Max=2 * 10
-8~2 * 10
-6Be fidelity iteration step length, μ
Min=μ
Max/ 100~μ
Max/ 1000 is slow iteration step length.
Table 1 is represented original neural net method, only the combination property by the original neural net method after this method pre-treatment step, the inventive method compares.
The comprehensive comparison of four kinds of algorithms of table 1.
Compare index | Original neural net method | Based on the motion determination neural net method of getting threshold value | Only by original neural net method after the pre-treatment step | The inventive method |
(Fig. 5 a-Fig. 8 a) for target fade-out | Target fade-out is arranged | Driftlessness is degenerated | Target fade-out is arranged | Driftlessness is degenerated |
Puppet resembles (Fig. 5 b-Fig. 8 b) (Fig. 5 c-Fig. 8 c) | There is puppet to resemble | No puppet resembles | There is puppet to resemble | No puppet resembles |
(Fig. 5 a-Fig. 8 a) for the little target signal to noise ratio of single frames (db) | 2.118986 | 7.106184 | 6.294982 | 11.524273 |
The degree of roughness of image (Fig. 5 c-Fig. 8 c) | 0.043690 | 0.055771 | 0.034379 | 0.034816 |
Fixed pattern pollutes the removal situation | And do not remove | And do not remove | Remove | Remove |
The method adaptivity | Need not artificially participate in | Need artificial constantly the adjustment to judge the threshold value of whether moving | Need not artificially participate in | Need not artificially participate in |
The little target signal to noise ratio of single frames shown in the table 1 (db) adopts following formula to calculate:
S wherein is the brightness average of target; μ
0Be background mean value, σ
0Be the background standard deviation; The signal to noise ratio of the little target of the big more representative of SNR value is high more, and effect is good more.As seen from Table 1, the method that the present invention proposes, the little target signal to noise ratio of single frames (db) maximum, be almost 5 times of original neural method, for 2 times by original neural net method after the pre-treatment step only, be 1.5 times based on the motion determination neural net method of getting threshold value.Calculate traditional two-point calibration method (driftlessness degeneration) of learning that infrared focal plane asymmetric is proofreaied and correct in addition, the little target signal to noise ratio of its single frames is 11.619666 under identical image and normalization situation, the little target signal to noise ratio of the inventive method single frames is suitable with it, so the inventive method can reach the purpose that prevents little target fade-out.
The degree of roughness of image adopts following formula to calculate:
Wherein, f represents digital image, h
1=[1 ,-1] is horizontal shuttering (horizontal mask),
Be vertical formwork (vertical mask) that * represents discrete convolution, ‖ ‖
1Represent L
1Norm.
This index can be polluted the performance that roughness visual and correction back image is come evaluation algorithms by calculating steady noise under the realistic images condition of unknown.Its value is the smaller the better.Demonstrated the pre-treatment step of use this paper by data result shown in the table 1 after, the original neural net method of roughness ratio of image and all low based on the motion determination neural net method of getting threshold value, the steady noise of presentation image pollute and are lowered.Visual roughness is also than original neural net method and all low based on the motion determination neural net method of getting threshold value, so calibration result is better after the correction that the method for this paper obtains.
Except the data target that calculates, the method (original neural net method and only by original neural net method after the pre-treatment step) that does not add exercise factor in the subjective assessment all exists target fade-out and artifact problem; Secondly, original neural net method and pollute based on the fixed pattern that the motion determination neural net method of getting threshold value is not all removed infrared focus plane and presented, especially proposed in the original document based on the motion determination neural net method of getting threshold value, need artificial the participation to adjust the motion determination threshold value, reduced the adaptivity of system.And method of the present invention has not only been removed the fixed pattern pollution, target fade-out and pseudomorphism have not occurred, also need not artificially participate in selecting threshold value and as the scale operation of traditional approach, has been improved the adaptivity of Nonuniformity Correction.
The present invention adopts that background frames is handled, bad unit detects and replaces and based on the method that the motion detection self adaptation is adjusted the neural net iteration step length, instructed neural net to proofread and correct, and effectively solves and the target fade-out and the puppet that have suppressed to produce in the scene class algorithm resemble phenomenon.Whether the motion that the target fade-out and the basic origin cause of formation that produces pseudomorphism are scene just should proofread and correct according to scene information when scene is constantly moved, otherwise should stop the renewal of scene information.So this method according to the motion variance of scene, as the direct ratio information of neural net iteration step length, when scene motion is abundant, increases iteration step length adaptively, when scene motion is slow, reduces iteration step length, thereby has controlled the renewal speed of correction coefficient.Like this motion sufficient the time correction coefficient can normally upgrade, slow down the renewal of correction coefficient and move when mild, even when scene is static, stop the renewal of correction coefficient.And this method uses pre-treatment step to remove fixed pattern noise before processing, do not need to set special motion detection threshold and instruct coefficient update, do not have complicated calculating operation (as division) yet, solved all weak points of existing method, be beneficial to hardware and realize, reach real-time.
Description of drawings
Fig. 1 is a schematic flow sheet of the present invention;
Fig. 2 is a bidirectional linear extrapolation operator pixel map, is used to judge bad unit;
Fig. 3 a~Fig. 3 c is respectively the background frames image of Fig. 4 a~Fig. 4 c;
Fig. 4 a, Fig. 4 b are the original infrared image (with black arrow and the little target location of white box indication aircraft, not belonging to original infrared imagery on the figure) of the little target of aircraft;
Fig. 4 c is the original infrared image of hand and mobile phone target;
(Fig. 5 a shows that with black arrow and the little target location of white box indication aircraft it the target fade-out phenomenon occurs to figure as a result after Fig. 5 a~Fig. 5 c is respectively and with original neural net method Fig. 4 a~Fig. 4 c is handled; Fig. 5 b shows that with black arrow and the little target location of white box indication aircraft it puppet occurs and resembles; The position of pseudo-elephant appears in Fig. 5 c with black arrow indication hand);
Fig. 6 a~Fig. 6 c is respectively the figure as a result that uses after based on the motion determination neural net method of getting threshold value Fig. 4 a~Fig. 4 c being handled (with black arrow and the little target location of white box indication aircraft, indicated with the white dashed line square frame does not have the processed first example of falling of evil idea to Fig. 6 a with black arrow and the little target location of white box indication aircraft, Fig. 6 b);
(Fig. 7 a shows that with black arrow and the little target location of white box indication aircraft it the target fade-out phenomenon occurs to figure as a result after Fig. 7 a~Fig. 7 c only is respectively and by the original neural net method after the pre-treatment step of the present invention Fig. 4 a~Fig. 4 c is handled; Fig. 7 b shows that with black arrow and the little target location of white box indication aircraft it puppet occurs and resembles; The position of pseudo-elephant appears in Fig. 7 c with black arrow indication hand);
(Fig. 8 a shows that with black arrow and the little target location of white box indication aircraft it the target fade-out phenomenon do not occur to figure as a result after Fig. 8 a~Fig. 8 c is respectively and by the present invention Fig. 4 a~Fig. 4 c is handled; Fig. 8 b shows that with black arrow and the little target location of white box indication aircraft it puppet do not occur and resembles; The position of pseudo-elephant does not appear in Fig. 8 c with black arrow indication hand).
Embodiment
Several groups of sequence images that collected from actual 128 * 128 specification LW MCT IRFPA have been adopted in following explanation, and acquisition frame rate was 100 frame/seconds.As shown in Figure 1, specific implementation process of the present invention is:
(1) pre-treatment step, before proofreading and correct in real time, cover the infrared focal plane detector lens cap or infrared focal plane detector is not had the blank local of scene facing to driftlessness, respectively the time domain average of two groups that under blank scene, collect each little target images of 10 frame aircrafts and one group of 10 frame hand and handset image as three groups of background frames image B, as Fig. 3 a~Fig. 3 c; Gather two groups of original infrared images of the little target of aircraft respectively with detector, as Fig. 4 a, the original infrared image of Fig. 4 b and one group of hand and mobile phone is as Fig. 4 c.Utilization is carried out invalid picture dot detection based on the invalid picture dot detection algorithm of scene to background frames, and testing result is stored among the bad meta template badTemplet;
Calculate the average gray B of effective picture dot (the first point of non-evil idea) in the background frames image B
AveThe gain correction coefficient of initialization first each picture dot of two field picture
Invalid picture dot (bad first point) detects and adopts following method to determine:
The image of the focal plane array gathered adopted 5 * 5 array carry out the calculating of bidirectional linear extrapolation operator, as shown in Figure 2 as kernel.Calculate the one dimension operator of two dimensional image in X, Y direction at first respectively, computing formula is as follows:
bad
xmax=max{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1}
bad
xmin=min{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1}
bad
ymax=max{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j}
bad
ymin=min{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j}
If the actual B that obtains
I, jGray value satisfy:
(B
I, j-bad
Xmax)/B
I, j>10% or (B
I, j-bad
Ymax)/B
I, j>10%
Judge that then this pixel is overheated invalid picture dot;
If the actual B that obtains
I, jGray value satisfy:
(bad
Xmin-B
I, j)/B
I, j>10% or (bad
Ymin-B
I, j)/B
I, j>10%
Judge that then this pixel was cold invalid picture dot.
(2) aligning step, input is as the frame original image V among Fig. 4 a~Fig. 4 c
n, its each picture dot point gray value V
I, j n, with V
I, j nThe gray value B of subtracting background frame correspondence position
I, jObtain this frame image X to be corrected
n, its each picture dot point gray value X
I, j n, subscript i wherein, j is respectively abscissa and the ordinate of picture dot in image, that is:
Read bad meta template badTemplet, effective picture dot (the first point of non-evil idea) is proofreaied and correct the image Y after obtaining proofreading and correct
n, its each picture dot point gray value Y
I, j n, updating formula is:
Wherein, G
I, j nBe this frame gain correction coefficient.
Pixel value to invalid picture dot position substitutes with the average of left neighborhood territory pixel and last neighborhood territory pixel sum simultaneously, and the image after obtaining at last proofreading and correct is as Fig. 8 a~Fig. 8 c.
(3) iteration step length set-up procedure, two frame corresponding points motion variances sigma before and after calculating
T, ij 2, the image corresponding points motion variances sigma after front and back two frames are proofreaied and correct
T, ij 2Adopt following method to determine:
T wherein
I, j nBe this frame coordinate i, the value behind the j picture dot point calibration, Y
I, j N-1Be former frame coordinate i, the value behind the j picture dot point calibration,
Be two two field picture coordinate i, the gray average of j picture dot point.
Judge σ
T, ij 2Whether>4000 set up, and is, then initialization once more
Change step (2); , then do not adjust iteration step length μ
I, j n, change step (4); Wherein 4000 be motion variance threshold values σ
Tmax 2
Adjust iteration step length μ
I, j nAdopt following method to determine:
Interim iteration step length
True iteration step length is as the iteration step length of final gain coefficient update, i.e. μ if interim iteration step length greater than the fidelity iteration step length, is then gone bail for
I, j'>μ
Max, then
Otherwise, i.e. μ
I, j'≤μ
Max, then with the iteration step length of the interim iteration step length of calculating gained as the final gain coefficient update,
Wherein the fidelity iteration step length represents to use neural net method timing, can guarantee the rapid convergence of correction parameter and don't can make visual correction of a final proof result exceed the iteration speed of effective range, can specifically be taken as μ
Max=2 * 10
-7, slowly iteration step length that is to say the iteration speed when correction parameter is slowly restrained, and can specifically be taken as μ
Min=μ
Max/ 100.
(4) gain correction coefficient step of updating, gain correction coefficient more new formula are:
F wherein
I, j nBe the image Y after proofreading and correct
I, j nFour field averages, f
I, j nComputing formula is:
Figure as a result after Fig. 5 a~Fig. 5 c is depicted as and with original neural net method Fig. 4 a~Fig. 4 c is handled; Fig. 6 a~Fig. 6 c is depicted as the figure as a result that uses after based on the motion determination neural net method of getting threshold value Fig. 4 a~Fig. 4 c being handled; Fig. 7 a~Fig. 7 c is depicted as the figure as a result after only by the original neural net method after this method pre-treatment step Fig. 4 a~Fig. 4 c being handled.Last three picture groups are compared with Fig. 8 a~Fig. 8 c, can obviously find, the figure effect that original neural net method is handled is also bad, just like among Fig. 5 a shown in the labeling position significantly the target fade-out problem and shown in labeling position among Fig. 5 b, Fig. 5 c significantly puppet resemble problem, and the fixed pattern that this method is not removed infrared focal plane imaging pollutes; Though use based on the motion determination neural net method of getting threshold value owing to added motion determination, prevented that little target fade-out and puppet from resembling problem, but it still pollutes just like the fixed pattern shown in Fig. 6 a~Fig. 6 c, and the problems such as bad unit shown in frame of broken lines among Fig. 6 b, and its not homotactic motion determination threshold value difference needs artificially selected; And the figure that only handles by the original neural net method after this method pre-treatment step resembles problem just like the target fade-out problem of Fig. 7 a with as the tangible puppet of Fig. 7 b and Fig. 7 c.Image after handling by this method be Fig. 8 a~Fig. 8 c, and as seen this method not only has adaptivity, and has effectively removed puppet and resemble, and has solved the target fade-out problem.
Claims (3)
1. infrared focal plane asymmetric correction method based on motion detection guidance, order comprises:
(1) pre-treatment step, the image with the even irradiation of infrared focal plane detector collection M frame carries out time domain average to them then and obtains the background frames image B, its each picture dot point gray value B
I, jUtilize the invalid picture dot detection algorithm based on scene to carry out invalid picture dot detection to each picture dot point of background frames image, testing result deposits bad meta template in, calculates the average gray B of effective picture dot in the background frames image according to testing result
AveSetting original sequence totalframes to be corrected is N, and N is a natural number, the gain correction coefficient of initialization each picture dot of original image to be corrected
Initialization original sequence n=1 to be corrected, M are 10~100;
(2) aligning step is imported n frame original image V
n, its each picture dot point gray value is V
I, j n, with V
I, j nThe gray value B of subtracting background frame correspondence position
I, j, obtain n frame image X to be corrected
n, its each picture dot point gray value is X
I, j n:
Subscript i wherein, j is respectively abscissa and the ordinate of picture dot in image;
Read bad meta template, treat correcting image X
nThe picture dot of effective picture dot position is proofreaied and correct in the corresponding bad meta template, and updating formula is:
Wherein, G
I, j nIt is n frame gain correction coefficient; Treat correcting image X simultaneously
nThe pixel value of corresponding invalid picture dot position substitutes with the mean value of left neighborhood picture dot pixel and last field picture dot pixel sum, obtains proofreading and correct back image Y at last
n, its each picture dot gray value is Y
I, j n
(3) iteration step length set-up procedure judges whether n>1, otherwise changes step (4); Be that two frames are proofreaied and correct the corresponding picture dot motion of back image variances sigma before and after then calculating
T, ij 2, judge whether
Otherwise adjust iteration step length μ
I, j n, change step (4), be then initialization once more
Original sequence n to be corrected adds 1, judges whether n>N, is then to finish, otherwise changes step (2), wherein motion variance threshold values σ
Tmax 2Be (0.5~1) B
Ave
(4) gain correction coefficient step of updating utilizes the neural net correcting algorithm that the gain heterogeneity of image is compensated, and promptly upgrades gain correction coefficient, and formula is:
F wherein
I, j nFor proofreading and correct back image Y
nIn the neighbours territory picture dot gray value average of each picture dot, after gain correction coefficient upgraded, original sequence n to be corrected added 1, judges whether n>N, is then to finish, otherwise changeed step (2).
2. the infrared focal plane asymmetric correction method based on motion detection guidance as claimed in claim 1 is characterized in that, in the described pre-treatment step, invalid picture dot testing process is:
(1) calculate the one dimension operator of background frames image in X, Y direction respectively, computing formula is:
bad
xmax=max{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1},
bad
xmin=min{2B
i,j-1-B
i,j-2,2B
i,j+1-B
i,j+2,B
i,j-1,B
i,j+1},
bad
ymax=max{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j},
bad
ymin=min{2B
i-1,j-B
i-2,j,2B
i+1,j-B
i+2,j,B
i-1,j,B
i+1,j};
(2) each picture dot point of background frames image is judged:
(B
i,j-bad
xmax)/B
i,j>badTH、(B
i,j-bad
ymax)/B
i,j>badTH、
(bad
Xmin-B
I, j)/B
I, j>badTH or (bad
Ymin-B
I, j)/B
I, j>badTH,
This picture dot of background frames image B B is then judged in arbitrary establishment in four formulas
I, jBe invalid picture dot, otherwise be effective picture dot.
3. the infrared focal plane asymmetric correction method based on motion detection guidance as claimed in claim 1 is characterized in that, in the described iteration step length set-up procedure:
(1) two frames are proofreaied and correct the corresponding picture dot motion of back image variances sigma before and after
T, ij 2For:
Y wherein
I, j N-1Be that the n-1 frame is proofreaied and correct back image coordinate i, the gray value of j picture dot,
Be that two frames are proofreaied and correct back image coordinate i, the gray average of j picture dot;
(2) adjust iteration step length μ
I, j nProcess is:
Make interim iteration step length
If μ
I, j'>μ
Max, then
If μ
I, j'≤μ
Max, then
μ wherein
Max=2 * 10
-8~2 * 10
-6Be fidelity iteration step length, μ
Min=μ
Max/ 100~μ
Max/ 1000 is slow iteration step length.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CNB2007100519206A CN100433793C (en) | 2007-04-19 | 2007-04-19 | Infrared focal plane asymmetric correction method based on the motion detection guidance |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CNB2007100519206A CN100433793C (en) | 2007-04-19 | 2007-04-19 | Infrared focal plane asymmetric correction method based on the motion detection guidance |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101056353A true CN101056353A (en) | 2007-10-17 |
CN100433793C CN100433793C (en) | 2008-11-12 |
Family
ID=38795946
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CNB2007100519206A Expired - Fee Related CN100433793C (en) | 2007-04-19 | 2007-04-19 | Infrared focal plane asymmetric correction method based on the motion detection guidance |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN100433793C (en) |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102214354A (en) * | 2011-04-29 | 2011-10-12 | 中国航空工业集团公司洛阳电光设备研究所 | Method for detecting invalid pixels of infrared image in real time |
CN101666682B (en) * | 2009-08-06 | 2012-05-30 | 重庆邮电大学 | Neural network nonuniformity correction method based on scene statistics |
CN103180718A (en) * | 2010-10-29 | 2013-06-26 | 奥林巴斯株式会社 | Image analysis method and image analysis device |
CN103218792A (en) * | 2013-05-03 | 2013-07-24 | 中国电子科技集团公司第二十八研究所 | Infrared image noise time domain filtering method based on registration |
CN104599261A (en) * | 2014-11-19 | 2015-05-06 | 北京泰和云信科技有限公司 | Equalization histogram neural network heterogeneity correcting method for short wave infrared focal plane |
CN105005967A (en) * | 2015-05-28 | 2015-10-28 | 西安电子科技大学 | Method and apparatus for correcting non-uniformity of infrared imaging based on combined space-time filtering |
CN105160657A (en) * | 2015-08-05 | 2015-12-16 | 西安电子科技大学 | Infrared imaging non-uniformity correction system based on FPGA |
CN106017695A (en) * | 2016-07-20 | 2016-10-12 | 上海航天控制技术研究所 | Adaptive infrared non-uniformity calibration method based on motion state estimation |
CN104166960B (en) * | 2014-07-24 | 2017-02-15 | 中国科学院光电技术研究所 | Self-adaptive non-uniform fixed noise removing method based on scene |
CN109272520A (en) * | 2018-09-18 | 2019-01-25 | 浙江大学 | A kind of adaptive Non-uniformity Correction of Infrared Focal Plane Arrays method of Union Movement guidance and edge detection |
CN109697700A (en) * | 2018-05-07 | 2019-04-30 | 杭州新瀚光电科技有限公司 | Hand-held thermal infrared imager shoots ant-shaking algorithm |
CN110428466A (en) * | 2019-07-16 | 2019-11-08 | 浙江大华技术股份有限公司 | A kind of method and apparatus of Nonuniformity Correction |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US4975864A (en) * | 1989-01-26 | 1990-12-04 | Hughes Aircraft Company | Scene based nonuniformity compensation for starting focal plane arrays |
US6507018B2 (en) * | 1996-08-30 | 2003-01-14 | Raytheon Company | Ditherless non-uniformity compensation for infrared detector arrays with recursive spatial low pass filtering |
US7463753B2 (en) * | 2004-09-15 | 2008-12-09 | Raytheon Company | FLIR-to-missile boresight correlation and non-uniformity compensation of the missile seeker |
CN1811360A (en) * | 2006-02-24 | 2006-08-02 | 昆明物理研究所 | Adaptive non-uniform correcting method for stare infrared focal plane detector |
CN100432697C (en) * | 2006-04-05 | 2008-11-12 | 中国科学院上海技术物理研究所 | Linear array scan image non-uniform correction method based on statistical restraint model |
-
2007
- 2007-04-19 CN CNB2007100519206A patent/CN100433793C/en not_active Expired - Fee Related
Cited By (17)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101666682B (en) * | 2009-08-06 | 2012-05-30 | 重庆邮电大学 | Neural network nonuniformity correction method based on scene statistics |
CN103180718B (en) * | 2010-10-29 | 2016-01-13 | 奥林巴斯株式会社 | Image analysis method and image analysis apparatus |
CN103180718A (en) * | 2010-10-29 | 2013-06-26 | 奥林巴斯株式会社 | Image analysis method and image analysis device |
CN102214354A (en) * | 2011-04-29 | 2011-10-12 | 中国航空工业集团公司洛阳电光设备研究所 | Method for detecting invalid pixels of infrared image in real time |
CN103218792A (en) * | 2013-05-03 | 2013-07-24 | 中国电子科技集团公司第二十八研究所 | Infrared image noise time domain filtering method based on registration |
CN103218792B (en) * | 2013-05-03 | 2015-07-08 | 中国电子科技集团公司第二十八研究所 | Infrared image noise time domain filtering method based on registration |
CN104166960B (en) * | 2014-07-24 | 2017-02-15 | 中国科学院光电技术研究所 | Self-adaptive non-uniform fixed noise removing method based on scene |
CN104599261A (en) * | 2014-11-19 | 2015-05-06 | 北京泰和云信科技有限公司 | Equalization histogram neural network heterogeneity correcting method for short wave infrared focal plane |
CN105005967A (en) * | 2015-05-28 | 2015-10-28 | 西安电子科技大学 | Method and apparatus for correcting non-uniformity of infrared imaging based on combined space-time filtering |
CN105160657A (en) * | 2015-08-05 | 2015-12-16 | 西安电子科技大学 | Infrared imaging non-uniformity correction system based on FPGA |
CN105160657B (en) * | 2015-08-05 | 2018-01-30 | 西安电子科技大学 | Infrared imaging Nonuniformity Correction system based on FPGA |
CN106017695A (en) * | 2016-07-20 | 2016-10-12 | 上海航天控制技术研究所 | Adaptive infrared non-uniformity calibration method based on motion state estimation |
CN106017695B (en) * | 2016-07-20 | 2019-02-19 | 上海航天控制技术研究所 | Adaptive infrared asymmetric correction method based on state estimation |
CN109697700A (en) * | 2018-05-07 | 2019-04-30 | 杭州新瀚光电科技有限公司 | Hand-held thermal infrared imager shoots ant-shaking algorithm |
CN109272520A (en) * | 2018-09-18 | 2019-01-25 | 浙江大学 | A kind of adaptive Non-uniformity Correction of Infrared Focal Plane Arrays method of Union Movement guidance and edge detection |
CN110428466A (en) * | 2019-07-16 | 2019-11-08 | 浙江大华技术股份有限公司 | A kind of method and apparatus of Nonuniformity Correction |
CN110428466B (en) * | 2019-07-16 | 2022-02-18 | 浙江大华技术股份有限公司 | Method and equipment for correcting nonuniformity |
Also Published As
Publication number | Publication date |
---|---|
CN100433793C (en) | 2008-11-12 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN101056353A (en) | Infrared focal plane asymmetric correction method based on the motion detection guidance | |
CN101038209A (en) | Infrared focal plane array heterogeneity self-adaptive correction method | |
CN106157267B (en) | Image defogging transmissivity optimization method based on dark channel prior | |
CN109801343B (en) | Annular artifact correction method based on images before and after reconstruction and CT control system | |
CN110807738B (en) | Blurred image non-blind restoration method based on edge image block sharpening | |
WO2019127059A1 (en) | Infrared image non-uniformity correction method employing guided filtering and high-pass filtering | |
CN114998141B (en) | Space environment high dynamic range imaging method based on multi-branch network | |
CN105654430A (en) | Contrast constrained pneumatic thermal radiation correction method | |
CN111598788B (en) | Single image defogging method based on quadtree decomposition and non-local prior | |
CN112614063B (en) | Image enhancement and noise self-adaptive removal method for low-illumination environment in building | |
CN1900971A (en) | Improved NAS RIF blind image recovery method | |
CN103868601B (en) | The bilateral full variational regularization bearing calibration of the non-homogeneous response of IRFPA detector | |
CN111932478A (en) | Self-adaptive non-uniform correction method for uncooled infrared focal plane | |
CN102831573B (en) | Linear stretching method of infrared image | |
CN105005967A (en) | Method and apparatus for correcting non-uniformity of infrared imaging based on combined space-time filtering | |
CN109615590B (en) | Sonar image enhancement method based on fuzzy algorithm and fractional differential algorithm | |
CN109191405B (en) | Aerial image defogging algorithm based on transmittance global estimation | |
CN113487516B (en) | Defogging processing method for image data | |
CN111028156A (en) | Mixed filtering image defogging algorithm based on dark channel prior model | |
Li et al. | A fast image dehazing algorithm for highway tunnel based on artificial multi-exposure image fusion | |
CN108805831B (en) | Image enhancement method under deep sea environment | |
CN111598815A (en) | Thermal infrared imager image non-uniformity correction method and device | |
Qian | An improved restoration algorithm for blurred images based on complete blind convolution and non-local means filtering | |
CN113313670B (en) | Underwater illumination non-uniform image enhancement method based on alternate direction multiplier method | |
CN117788344B (en) | Building texture image restoration method based on diffusion model |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20081112 Termination date: 20170419 |