CN1312636C - Morphologic filter automatic destination detecting method - Google Patents

Abstract

The present invention relates to a method for detecting an automatic destination of a morphologic filter. Firstly, training samples used for optimizing training structure elements are collected, the samples can comprise various point destinations and backgrounds as much as possible to form a genetic algorithm used for optimizing training work, the genetic algorithm adopts novel interval discretization codes and self-adaptive major-minor cross and mutation operators, the values of the training structure elements are optimized by the genetic algorithm through the collected samples, a morphologic filter based on Top-Hat operators is formed on the base of the optimized structure elements to carry out wave filtration to infrared destination images, finally, the majority small detected point destinations are divided according to a self-adaptive threshold, and the point destinations with high signal-noise rate are divided by a fixed threshold to detect destination points. The present invention realizes the automatic detection for infrared small point destinations under the condition of a complex background, the destination detecting probability and the interference rejection are greatly enhanced, and the present invention has the extremely wideapplication foreground on civil and military aspect.

Description

Morphologic filter automatic destination detecting method

Technical field

The present invention relates to a kind of object detection method that is used for technical field of image processing, specifically is a kind of morphologic filter automatic destination detecting method.

Background technology

In recent years, along with to morphologic research and development, this special image processing subject of morphology image processing develops into a main research field of image processing gradually, and progressively becomes the favourable instrument of Weak target detection and Identification.The morphologic filtering device can be decomposed into morphology operations and these two basic problems of structural element.Corrosion is expanded, and the open and close operator is 4 kinds of basic operators of morphology operations, these 4 basic operators is made up can draw the morphological operator with different qualities.After the morphology operations rule was determined, the final filtering performance of morphological filter just only depended on the selection of structural elements, comprised the shape and the element value of structural element.Select different structural elements can cause analysis and the processing of computing to different geometry information.About utilizing in the research that morphological operator detects infrared small object, structural element all was to have determined in advance in the past.Therefore, these wave filters only have filtering performance preferably in pairing certain class iconic model.Yet generally picture signal is very complicated and be among the continuous variation, and this structural element that just requires to select for use should have adaptation function, to realize optimal treatment.

At present, for the optimization training of structural element, the researcher has proposed morphology neural network and two kinds of learning methods of morphology genetic algorithm both at home and abroad.Wherein, genetic algorithm is combined with the morphologic filtering device, utilize genetic algorithm training morphologic filtering device structural element to realize that optimal treatment is a kind of method of Chinese scholars primary study in infrared point target detection problem in recent years.

Find through literature search prior art, people such as Terebes R are at " Signal Processing, 20026th International Conference on " (Volume:1,26-30 Aug.2002.Pages:853-857vol.1) " the Adaptive Filtering UsingMorphological Operators and Genetic Algorithms " that delivers in (" the 6th signal Processing international conference in 2002 "), the intersection and the mutation operator of the genetic algorithms use routine that proposes in (" based on the auto adapted filtering of morphological operator and genetic algorithm ") this article make that its efficient in the convergence searching process is not high.Genetic algorithm is to find global optimum's point on probability meaning, institute's data volume to be processed is big, simultaneously, for as the first-selection that detects infrared small point target, also comparatively weak for the adaptive optimization training research of Top-Hat morphologic filtering device structural element at present.

Summary of the invention

The objective of the invention is to overcome the deficiency that prior art exists, a kind of morphologic filter automatic destination detecting method is proposed, it is trained based on genetic algorithm optimization, adopt the genetic algorithm of interval discretize coding and adaptive primary and secondary formula intersection and mutation operator, to be used for optimizing the training structure element, simultaneously, select Top-Hat high-pass filtering operator as morphologic filtering device operator, thereby it is not high effectively to overcome existing genetic algorithm optimization morphologic filtering device technology median filter optimal performance, the genetic algorithm converges time is longer, the not high shortcoming of optimizing efficient.

The present invention is achieved by the following technical solutions, at first gather the training sample that is used for optimizing the training structure element, these samples should comprise various point targets and background as far as possible, be configured to optimize the genetic algorithm of training, interval discretize coding and adaptive primary and secondary formula that this genetic algorithms use is new are intersected and mutation operator, the sample that utilization collects is with this genetic algorithm optimization training structure element value, optimize at these that structure carries out filtering based on the morphologic filtering device of Top-Hat operator to the infrared target image on basis of good structural element.At last adopt adaptive thresholds to cut apart, the higher point target of signal to noise ratio is cut apart with fixed threshold detected impact point at the most of small point targets that detected.

Below to the further instruction of the present invention, comprise the steps:

(1) is configured to optimize the genetic algorithm of training

Specifically comprise coding, the ideal adaptation degree calculates, and intersects and makes a variation these several steps.Wherein, coded system is taked a kind of new interval discretize coding, and this is equivalent to the value real number scope discretize with individual component.Calculate each chromosomal adaptation functional value in the colony then, constitute a population with probability shown in the genetic algorithm STEP3 hereinafter with wheel disc back-and-forth method some chromosomes of picked at random from colony, intersect by adaptive primary and secondary formula at last and generate an individual new colony with the variation computing.

Genetic algorithm is colony's optimizing process, and it is to be set out by one group of initial value (biotic population) to be optimized.Here, colony is meant the finite point set of state space, represents with pop.The so-called process of optimizing is exactly that the pop of this colony constantly multiplies, competition, the process of heredity and variation.According to the biology term, with arbitrary finite point set { a among the pop ₁..., a _nBeing called a population, n is the scale of this population.Any point pop among the pop _i=a _i ⁽¹⁾a _i ⁽²⁾A _i ^(D)(1≤i≤n) is called body or chromosome one by one, and pop _iIn each a _i ^(j)(l≤j≤D) is referred to as gene.The genetic algorithm general description is as follows:

STEP1 selects a coding of problem; Provide one N chromosomal initial population pop (1), t=1 are arranged.

STEP2 is to each the chromosome pop among the pop of colony (t) _i(t) calculate its adaptive value function

f _i＝fitness(pop _i(t))

STEP3 is if stopping rule satisfies, and then algorithm stops; Otherwise, calculating probability

$P_i=\frac{\exp [f_i/T_k]}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\exp [f_j/T_k]},i=\mathrm{1,2},\…,N$

{ T wherein _kBe gradually in 0 annealing temperature, and $T_k=1/\ln [{\frac{k}{T}}_0+1],T_0=100,k=1,2,\·\·\·.$

And from pop (t), select some chromosomes to constitute a population at random with this probability distribution

newpop(t+1)＝{pop _j(t)|j＝1，2，…，N}

STEP4 is by mating, and the mating probability is P _c, obtaining one has N chromosomal crosspop (t+1).

STEP5 is with a less probability P _m, make chromosomal gene morph, form mutpop (t+1);

Make t=t+1, the new pop of colony (t)=mutpop (t) generates; Return STEP2.

(2) cutting apart based on adaptive threshold

Determining of thresholding should be at each n * n elementary area, employing single frames detection probability, and false-alarm probability and signal to noise ratio (S/N ratio) are decided thresholding.The detection that adaptive threshold is cut apart for small point target is very effective, but but can not well detect for the higher point target of signal to noise ratio (S/N ratio).This is because of the raising along with the point target signal to noise ratio (S/N ratio), and the speed that adaptive threshold increases is far longer than point target through opening the growth rate that is worth after the surplus computing.Make it to adapt if reduce the growth rate of adaptive threshold, can not effectively detect small point target again with the growth rate of target signal to noise ratio.For this reason, among the present invention at the given threshold value of the mean square deviation of each n * n elementary area, because the higher point target of signal to noise ratio (S/N ratio) utilizes fixed threshold just can well it be detected, utilize fixed threshold to cut apart detection so be higher than the point target of this threshold value, the small point target that is lower than this threshold value utilizes adaptive threshold to cut apart detection.

In the method for the present invention, morphological operator is selected the Top-Hat operator with high-pass filtering for use, can effectively improve wave filter to the detectability of impact point with to the inhibition ability of ground unrest.Be used to optimize new interval discretize coding and the adaptive intersection and the mutation operator of genetic algorithms use of training wave filter, it is not high effectively to have overcome the optimization performance that adopts conventional coding to be brought in the previous methods, shortcomings such as convergence time is long, and improve ageing in the genetic algorithm converges searching process, thereby the filtering performance of bigger raising morphologic filtering device.The present invention uses future widely having aspect the military affairs civilian two, can lay the foundation for the guidance precision that improves China's surface-to-air ballistic missile, the firing area (medium-long range) that enlarges guided missile, this technology also helps to improve the search early warning tracking performance of ground electronic support system simultaneously, improves the military equipment strength of China greatly.

Description of drawings

Fig. 1 is the morphology Top-Hat wave filter theory diagram that the present invention is based on the genetic algorithm optimization training

Fig. 2 carries out after the Filtering Processing comparison diagram as a result for the present invention to continuous four width of cloth low signal-to-noise ratio images.

Wherein, Fig. 2 (a) (b), (c), (d) is four continuous frame raw images, and Fig. 2 (e) (f), (g), (h) carries out filtered image as a result for utilization the present invention to them.

Fig. 3 is for using the comparison diagram after the present invention detects the small point target in the same images respectively with the Top-Hat wave filter of training based on Neural Network Optimization.

Wherein, Fig. 3 (a) is for treating the raw image of filtering, and Fig. 3 (b) carries out filtered image as a result for using the Top-Hat wave filter based on the Neural Network Optimization training to raw image, and Fig. 3 (c) carries out filtered image as a result for utilization the present invention to raw image.

Embodiment

In order to understand technical scheme of the present invention better, embodiments of the present invention are described further below in conjunction with accompanying drawing.

The theory diagram that the morphologic filtering device that the present invention is based on genetic algorithm optimization training detects infrared small point target as shown in Figure 1, filtering mainly is divided into morphologic filtering and thresholding is cut apart two parts.Wherein morphologic filtering is the emphasis of target detection, and the morphologic filtering device can be decomposed into morphology operations and these two basic problems of structural element.After the morphology operations rule was determined, the final filtering performance of morphological filter just only depended on the selection of structural elements.The present invention utilizes a series of sample datas that obtain in advance with genetic algorithm the filter construction element to be trained, to obtain best filter parameter.Wherein, genetic algorithm comprises coding, and the ideal adaptation degree calculates, and intersects and makes a variation these several steps.Raw image by the morphologic filtering device filtering after the genetic algorithm optimization training after, adopt adaptive thresholds to cut apart at the most of small point targets that detected at last, the higher point target of signal to noise ratio cut apart with fixed threshold detected impact point.

The concrete implementation detail of each several part is as follows:

1. be configured to optimize the genetic algorithm of training

Morphologic filtering device parameter mainly is made of each component value of structural element, and its training study process belongs to the multi-parameters optimization problem.For this reason, when they are mapped as in the hereditary space string structure data by genomic constitution, adopt the multiparameter coded system.The present invention introduces a kind of new interval discretize coding, supposes that each component value of individual B is interval [x _Min, x _Max] in all real numbers.At this moment, determine the string length K of structure earlier, then at [x _Min, x _Max] equal intervals ground insertion 2 ^K-2 points, every adjacent 2 distance is

$\mathrm{\δ}=\frac{x_{\max }-x_{\min }}{2^k-1}$

Then at [x _Min, x _Max] on chosen 2 ^KIndividual, they are respectively x _Min, x _Min+ δ, x _Min+ 2 δ ..., x _Min+ (2 ^K-1) δ=x _Max, these are 2 years old ^KIndividual point represents with the K bit respectively, promptly

Interval discretize coding and conventional coding are carried out emulation respectively, and shown in simulation result such as the table 3, table 4, table 3 compares for the present invention's algorithm convergence when adopting conventional coding genetic training consumes CPU time under the different fitness situations.Table 4 is comparing the ideal adaptation degree under certain cycle index when training with the conventional coding genetic of employing for the present invention.Table 3 has reflected that the interval discretize coding genetic speed of convergence among the present invention is greater than conventional coding genetic speed of convergence, and table 4 has reflected that its optimization performance is better than the optimization performance of conventional coding genetic.

Need to introduce relevant priori in the genetic algorithm learning rules and statistical law uses restraint, and provide preferred standard (cost function) with the guiding solution procedure.For the optimization of filtering parameter training, most critical be the expectation value that the Nonlinear Mapping output of morphological filter will approach training sample as best one can, promptly require optimum solution to be consistent and be optimum description with all examples.Simultaneously, also need take into account the navigability that stops algorithm (shutdown criterion).Therefore, the present invention selects for use the square error cost function of optimum solution target comparatively desirable with the cost function of traction optimization searching as correction, is defined as follows:

$E=\frac{1}{2L}\underset{k=1}{\overset{L}{\mathrm{\Σ}}}{(Y_k-d_k)}^2$

Here L is a number of training, d _kFor exporting the expectation value of k corresponding input signal, Y _kBe defined as the maximal value of Top-Hat morphologic filtering device at k training sample input back output matrix, as follows:

When opening surplus computing

Y _k＝max(F _k-(F _k·B)(x))

When closing surplus computing

Y _k＝max((F _k·B)(x)-F _k)

So individual pop _i(t) adaptive value function f _i=fitness (pop _i(t)) be defined as:

$f_i=1/E=2L/\underset{k=1}{\overset{L}{\mathrm{\Σ}}}{(Y_k-d_k)}^2$

Parents' chromosome produces child chromosome in the mode of intersecting under certain probability, thereby makes offspring individual heredity parents' essential characteristic.The present invention newly designs a kind of primary and secondary formula crossover operator, its thought is at the algorithm initial stage, and crossing operation is primarily aimed near the high weight component the initial point, to strengthen formation speed and the generating probability that high-quality is separated, quicken to eliminate poor quality and separate, have the solution space of optimizing potentiality thereby region of search is turned to early; In the algorithm later stage, crossing operation then emphasis is optimized good high weight component with protection, and is optimized less important component gradually at the low weight component at individual edge.For this reason, it is the rectangle at center with the initial point that the structural element of big or small n * n can be seen as a series of, and the big or small degree of its certain component weight is measured with the length of side of its place rectangle.Its length of side of the rectangle of i * i is defined as i in the structural element.

The individual pop of last 2 structural elements of given vector set Q ₁(t) and pop ₂(t). primary and secondary formula crossover probability is defined as follows:

$P_c=\left\{\begin{array}{cc}{K}_1\exp [-\mathrm{\&alpha;L}]\exp [-\stackrel{\&OverBar;}{f}/\mathrm{\&eta;}]& \stackrel{\&OverBar;}{f}<\mathrm{\&eta;}\\ K_2\exp \left[\mathrm{\&beta;L}\right]\exp [-\stackrel{\&OverBar;}{f}/\mathrm{\&eta;}]& \stackrel{\&OverBar;}{f}\&GreaterEqual;\mathrm{\&eta;}\end{array}\right.$

η is the fitness constant in the formula, and L is the length of side of individual component place rectangle, K ₁, K ₂Be constant, α, β＞0 is the weight constant, Average adaptive value for colony is defined as follows

$\stackrel{\&OverBar;}{f}=\sqrt{f_1^2+f_2^2+\&CenterDot;\&CenterDot;\&CenterDot;f_n^2}$

When each crossing operation carries out, with the probability shown in the genetic algorithm STEP3 with the wheel disc back-and-forth method from quantity be the colony of N a picked at random N/2 sample as population.It is regular as follows that the present invention takes a kind of new traversal formula to intersect:

{ crosspop _i, crosspop _(N/2)+i}=cop (newpop _i, newpop _s) 1≤i≤N/2 wherein, cop () is for intersecting function, newpop _iBe i individuality in the population, newpop _sThe individuality that from population, chooses with the wheel disc back-and-forth method with probability shown in the genetic algorithm STEP3 during for each crossing operation, crosspop _i, crosspop _N/2+iBe newly-generated individuality behind each crossing operation.Take turns after crossing operation finishes at each like this, the new colony chromosome number that produces is the same with former colony chromosome number.

Mutation operator is realized the optimization improvement of colony, for some gene that may lose in the intersection process are repaired and replenished, recovers the diversity that colony loses, to avoid being absorbed in local optimum.

The present invention according to the former head after time, the preferential optimization principles of weight is taked the variation of new primary and secondary formula, at the algorithm initial stage, mutation operation is primarily aimed near the high weight component the individual initial point, to reduce this stage to the unnecessary blind search of less important component; And be primarily aimed at the low weight component at individual edge at algorithm later stage mutation operation, keeping optimizing good fundamental component, and the less important component of progressive optimization.

Definition primary and secondary formula variation probability is as follows:

$P_m(L,t)=\left\{\begin{array}{cc}{P}_1\exp [-\frac{L^2}{{2\mathrm{\&sigma;}}^2}]& t\&le;\frac{T}{2}\\ P_2\exp \left[\frac{L^2}{{2\mathrm{\&sigma;}}^2}\right]\exp [-\frac{t}{\mathrm{\&tau;}}]& t>\frac{T}{2}\end{array}\right.$

In the formula, P ₁, P ₂Be the amplitude constant, σ is for regulating the constant of individual parameter weight size, and τ is for regulating P _mDecay the in time time constant of speed, T is the maximum algebraically that genetic algorithm is provided with.

2. cutting apart based on adaptive threshold

The present invention is cut apart this thought to the point target that signal to noise ratio is higher with fixed threshold according to adopting adaptive threshold to cut apart to the most of small point targets that detected, and adopts the single frames detection probability, and false-alarm probability and signal to noise ratio (S/N ratio) definition adaptive threshold are as follows:

$\mathrm{\&nu;}=\left\{\begin{array}{cc}u+\mathrm{SNR}-{\mathrm{\&sigma;\&Phi;}}^{-1}\left(P_d\right)& \mathrm{\&sigma;}\&le;t\\ T& \mathrm{\&sigma;}>t\end{array}\right.$

Wherein, p _dBe the single frames detection probability, SNR is a signal to noise ratio (S/N ratio), and v is a detection threshold, and u is the noise average after certain n * n image cell background offsets, σ ²Be the noise mean square deviation.U and σ ²Ask method as shown in the formula

$f=g-\left(\mathrm{g\&Theta;B}\right)\&CirclePlus;B$

$u=\frac{1}{9}\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}f(i,j)$

${\mathrm{\&sigma;}}^2=\frac{1}{9}\underset{i=1}{\overset{3}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{3}{\mathrm{\&Sigma;}}}{(f(i,j)-u)}^2$

G is the raw image gray scale, and f is the gray scale behind the shape filtering opening operation.

Utilization the present invention carries out Filtering Processing to a series of infrared small point target images, and the filtering result as shown in Figure 2.And carry out emulation respectively relatively with adopting different structure element wave filter and different shape operator wave filter, shown in result such as the table 1, table 2.Table 1 carries out the comparison of single frames Filtering Processing for the present invention and the other two kinds wave filters of taking the different structure element to the image of the different signal to noise ratio (S/N ratio)s of 100 width of cloth.Wherein, mode (1) is carried out Filtering Processing for the Top-Hat morphologic filtering device with the fixed sturcture element, mode (2) is for to have trained the Top-Hat morphologic filtering device of structural element to carry out Filtering Processing with Neural Network Optimization, and mode (3) is for to carry out Filtering Processing with the present invention.Table 2 carries out the single frames Filtering Processing relatively for the present invention with adopting different shape operator wave filter under different state of signal-to-noise.Method therefrom of the present invention as can be seen with adopt the fixed sturcture element, with Neural Network Optimization training structure element and adopt the wave filter of other morphological operator to compare, can significantly improve filtering performance to the small point target image.Think comparison with the morphologic filtering device that adopts Neural Network Optimization training structure element, the present invention can also detect some it can not detected small point target, as shown in Figure 3.

Contrast existing infrared small object detection technique, the present invention can effectively suppress various noise, and the point target lower to signal to noise ratio (S/N ratio) accurately detects, and can reach the practical function of through engineering approaches.Simultaneously from whole performing step as can be known, the inventive method is easy to realize, thereby the through engineering approaches that detects for infrared small point target provides a technology implementation method.

Table 1 different structure element filter detection destination probability contrast table

SNR	Mode (1) is the detection probability of point target down	Mode (2) is the detection probability of point target down	Mode (3) is the detection probability of point target down
				1.5	90％	95％	97％

2	94％	96％	98％
				5	100％	100％	100％

Table 2 different shape is learned operator filter detection destination probability contrast table

SNR	The detection probability of point target under the dilation operation	The detection probability of point target under the opening operation	The detection probability of point target under the Top-Hat computing
				2	81％	92％	98％
3	84％	95％	100％

Table 3 different coding mode genetic algorithm converges timetable

Fitness	Take CPU time during conventional the coding	Interval discretize takies CPU time when encoding
			0.3355	61.7720	23.2340
0.3550	114.8130	39.0310

Table 4 different coding mode genetic algorithm optimization training performance contrast table

Cycle index	Ideal adaptation degree during conventional the coding	Ideal adaptation degree when interval discretize is encoded
			10	0.3109	0.3637
20	0.3264	0.3659

Claims (3)

1, a kind of morphologic filter automatic destination detecting method, it is characterized in that, genetic algorithm and Top-Hat morphologic filtering device are organically combined, at first gather the training sample that is used for optimizing the training structure element, these samples should comprise various point targets and background as far as possible, be configured to optimize the genetic algorithm of training, interval discretize coding and adaptive primary and secondary formula that this genetic algorithms use is new are intersected and mutation operator, the sample that utilization collects is with this genetic algorithm optimization training structure element value, optimize on the basis of good structural element structure based on the morphologic filtering device of Top-Hat operator at these, the infrared target image is carried out filtering, carry out based on the cutting apart of adaptive threshold at last, the higher point target of signal to noise ratio is cut apart with fixed threshold detected impact point at the most of small point targets that detected;

Described new interval discretize coding is specially: each component value of supposing individual B is interval [x _Min, x _Max] in all real numbers, determine earlier the string length K of structure, then at [x _Min, x _Max] equal intervals ground insertion 2 ^K-2 points, every adjacent 2 distance is $\mathrm{\&delta;}=\frac{x_{\max }-x_{\min }}{2^K-1},$ Then at [x _Min, x _Max] on chosen 2 ^KIndividual, they are respectively x _Min, x _Min+ δ, x _Min+ 2 δ ..., x _Min+ (2 ^K-1) δ=x _Max, these are 2 years old ^KIndividual point represents with the K bit respectively, promptly

2. morphologic filter automatic destination detecting method according to claim 1 is characterized in that, the described genetic algorithm that is configured to optimize training is specially:

Comprise coding, the ideal adaptation degree calculates, and intersects and makes a variation these several steps, wherein, coded system is taked a kind of new interval discretize coding, and this is equivalent to the value real number scope discretize with individual component, calculate each chromosomal adaptation functional value in the colony then, with probability $p_1=\frac{\exp [f_1/T_k]}{\underset{J=1}{\overset{N}{\mathrm{\&Sigma;}}}\exp [f_J/T_k]}$ , i=1,2 ..., N constitutes a population with wheel disc back-and-forth method some chromosomes of picked at random from colony, intersects by adaptive primary and secondary formula at last to generate a new colony with the variation computing, wherein: f ₁It is each the chromosome adaptive value function in the colony; { T _kBe gradually in 0 annealing temperature, and k=1,2 ...

3, morphologic filter automatic destination detecting method according to claim 1 is characterized in that, and is described based on the cutting apart of adaptive threshold, and is specially:

According to the noise average after certain n * n image cell background offsets, the noise mean square deviation, single frames detection probability and snr computation go out the self-adaptation threshold value, then at the given threshold value of the mean square deviation of each n * n elementary area, most of small point targets that the noise mean square deviation is lower than this threshold value adopt adaptive threshold to cut apart, and the point target that is higher than this threshold value is cut apart with fixed threshold.

Images