CN1312636C - Morphologic filter automatic destination detecting method - Google Patents

Abstract

A morphological filter automatic target detection method, first collect training samples used to optimize the training structure elements, the samples should contain a variety of point targets and backgrounds as much as possible, and construct a genetic algorithm for optimal training, the genetic algorithm uses a new Interval discretization coding and adaptive primary and secondary crossover and mutation operators, using the collected samples to optimize the value of the training structural elements with the genetic algorithm, and constructing the Top-Hat operator based on these optimized structural elements The morphological filter filters the infrared target image, and finally performs segmentation based on an adaptive threshold for most of the detected weak point targets, and uses a fixed threshold to segment and detect target points for point targets with a high signal-to-noise ratio. The invention realizes the automatic detection of weak and small infrared point targets under complex background conditions, greatly improves the target detection probability and anti-interference ability, and has extremely wide application prospects in civil and military applications.

Description

Automatic object detection method based on morphological filter

technical field

The invention relates to a target detection method used in the technical field of image processing, in particular to a morphological filter automatic target detection method.

Background technique

In recent years, with the development of research on morphology, morphological image processing, a special subject of image processing, has gradually developed into a major research field of image processing, and has gradually become a useful tool for weak and small target detection and recognition. Morphological filters can be decomposed into two basic problems of morphological operations and structural elements. Erosion, dilation, opening and closing operators are the four basic operators of morphological operations. Combining these four basic operators can result in morphological operators with different characteristics. When the morphological operation rules are determined, the final filtering performance of the morphological filter only depends on the selection of structural elements, including the shape and element value of structural elements. The selection of different structural elements will lead to the analysis and processing of different geometric structure information. In previous researches on using morphological operators to detect small infrared targets, the structural elements were determined in advance. Therefore, these filters only have good filtering performance in a certain type of corresponding image model. However, under normal circumstances, image signals are extremely complex and are constantly changing, which requires that the selected structural elements should have adaptive functions to achieve optimal processing.

At present, for the optimization training of structural elements, researchers at home and abroad have proposed two learning methods: morphological neural network and morphological genetic algorithm. Among them, combining the genetic algorithm with the morphological filter and using the genetic algorithm to train the structural elements of the morphological filter to achieve optimal processing is a method that domestic and foreign scholars have focused on in the infrared point target detection problem in recent years.

Found through literature search to prior art, people such as Terebes R are in " Signal Processing, 20026th International Conference on " (Volume: 1,26-30 Aug.2002.Pages: 853-857vol.1) (" 2002 the 6th "Adaptive Filtering Using Morphological Operators and Genetic Algorithms" published at the Second International Conference on Signal Processing), ("Adaptive Filtering Based on Morphological Operators and Genetic Algorithms") The genetic algorithm proposed in this paper uses conventional crossover and mutation operators , making it inefficient in the process of convergence optimization. The genetic algorithm is to find the global optimal point in the sense of probability, and the amount of data to be processed is large. At the same time, as the first choice for detecting infrared weak point targets, the current research on adaptive optimization training of Top-Hat morphological filter structural elements is relatively limited. weak.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art, and propose a morphological filter automatic target detection method, which optimizes training based on genetic algorithms, adopts interval discretization coding and self-adaptive primary and secondary crossover and mutation operators The genetic algorithm is used to optimize the training structural elements. At the same time, the Top-Hat high-pass filter operator is selected as the morphological filter operator, so as to effectively overcome the optimal performance of the filter in the existing genetic algorithm optimization morphological filter technology. High, the convergence time of the genetic algorithm is long, and the optimization efficiency is not high.

The present invention is achieved through the following technical solutions. First, collect training samples for optimizing training structural elements. These samples should contain various point targets and backgrounds as much as possible, and construct a genetic algorithm for optimizing training. The genetic algorithm adopts a new Interval discretization coding and adaptive primary and secondary crossover and mutation operators, using the collected samples to optimize the value of the training structural elements with the genetic algorithm, and constructing the Top-Hat operator based on these optimized structural elements Morphological filter, which filters the infrared target image. Finally, the self-adaptive threshold is used to segment most of the detected weak point targets, and the fixed threshold is used to segment and detect target points for point targets with relatively high signal-to-noise ratio.

The following is the further description of the present invention, including the steps:

(1) Construct a genetic algorithm for optimal training

It specifically includes the steps of encoding, individual fitness calculation, crossover and mutation. Among them, the encoding method adopts a new interval discretization encoding, which is equivalent to discretizing the real number range of the value of the individual component. Then calculate the fitness function value of each chromosome in the population, use the roulette selection method to randomly select some chromosomes from the population to form a population with the probability shown in the genetic algorithm STEP3 below, and finally generate a population through adaptive primary and secondary crossover and mutation operations a new group.

Genetic algorithm is a group optimization process, which starts from a group of initial values (biological group) for optimization. Here, a population refers to a finite set of points in the state space, denoted by pop. The so-called optimization process is the process of continuous reproduction, competition, inheritance and mutation of this group of pops. According to biological terms, any finite point set {a ₁ ,..., a _n } in pop is called a population, and n is the size of the population. Any point in pop _i = a _i ⁽¹⁾ a _i ⁽²⁾ ...a _i ^(D) (1≤i≤n) is called an individual or chromosome, and each _{a i (} ^j) (l ≤j≤D) are called genes. Genetic algorithms are generally described as follows:

A coding of STEP1 selection problem; given an initial population pop(1) with N chromosomes, t=1.

STEP2 calculates its fitness value function for each chromosome pop _i (t) in the population pop(t)

f _i =fitness(pop _i (t))

STEP3 If the stopping rule is satisfied, the algorithm stops; otherwise, calculate the probability

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ]</span>}{\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ]</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&hellip;</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; N</span>}$

where {T _k } is the annealing temperature tending to 0, and $T_k=1/\ln [{\frac{k}{T}}_0+1],T_0=100,k=1,2,\&Center\; Dot;\&Center\; Dot;\&Center\; Dot;.$

And use this probability distribution to randomly select some chromosomes from pop(t) to form a population

newpop(t+1)={pop _j (t)|j=1, 2,..., N}

STEP4: Through mating, the mating probability is P _c , and a crosspop(t+1) with N chromosomes is obtained.

STEP5 mutates the genes of the chromosome with a small probability P _m to form mutpop(t+1);

Let t=t+1, a new group pop(t)=mutpop(t) is generated; return to STEP2.

(2) Segmentation based on adaptive threshold

The determination of the threshold should be based on the detection probability of a single frame, the false alarm probability and the signal-to-noise ratio for each n×n image unit. Adaptive threshold segmentation is very effective for the detection of small point targets, but it cannot detect point targets with high signal-to-noise ratio. This is because as the signal-to-noise ratio of the point target increases, the growth rate of the adaptive threshold is far greater than the growth rate of the value of the point target after the remainder operation. If the growth rate of the adaptive threshold is reduced to adapt to the growth rate of the target signal-to-noise ratio, weak point targets cannot be effectively detected. For this reason, in the present invention, a threshold is given for the mean square error of each n×n image unit, because point targets with a higher signal-to-noise ratio can be detected well by using a fixed threshold, so higher than this threshold Point targets are segmented and detected with a fixed threshold, and weak point targets below this threshold are segmented and detected with an adaptive threshold.

In the method of the present invention, the morphological operator selects the Top-Hat operator with high-pass filter performance, which can effectively improve the detection ability of the filter to the target point and the suppression ability to the background noise. The genetic algorithm used to optimize the training filter adopts new interval discretization coding and adaptive crossover and mutation operators, which effectively overcomes the disadvantages of low optimization performance and long convergence time caused by conventional coding in previous methods , and improve the timeliness of the genetic algorithm convergence optimization process, thereby greatly improving the filtering performance of the morphological filter. The present invention has wide application prospects in both military and civilian uses, and can lay the foundation for improving the guidance accuracy of my country's surface-to-air missiles and expanding the attack range (middle and long range) of missiles. The early warning and tracking performance greatly improves the military equipment strength of our country.

Description of drawings

Fig. 1 is the principle block diagram of the morphology Top-Hat filter based on genetic algorithm optimization training of the present invention

Fig. 2 is a comparison diagram of the results after filter processing of four consecutive low SNR images according to the present invention.

Wherein, Fig. 2 (a), (b), (c), (d) are continuous four frames of original images, and Fig. 2 (e), (f), (g), (h) is to use the present invention to They are the resultant image after filtering.

Fig. 3 is a comparison diagram after the application of the present invention and the Top-Hat filter based on neural network optimization training to respectively detect weak and small point targets in the same image.

Among them, Fig. 3 (a) is the original image to be filtered, Fig. 3 (b) is the result image after filtering the original image with the Top-Hat filter based on neural network optimization training, Fig. 3 (c) It is the result image after filtering the original image by using the present invention.

Detailed ways

In order to better understand the technical solutions of the present invention, the implementation manners of the present invention will be further described below in conjunction with the accompanying drawings.

The principle block diagram of the detection of weak and small infrared point targets by the morphological filter based on genetic algorithm optimization training of the present invention is shown in Fig. 1, and the filtering process is mainly divided into two parts: morphological filtering and threshold segmentation. Among them, morphological filtering is the focus of target detection, and morphological filtering can be decomposed into two basic problems of morphological operation and structural elements. When the morphological operation rules are determined, the final filtering performance of the morphological filter only depends on the selection of structural elements. The invention utilizes a series of sample data obtained in advance to train the filter structural elements with a genetic algorithm to obtain the best filter parameters. Among them, the genetic algorithm includes coding, individual fitness calculation, crossover and mutation. After the original image is filtered by the morphological filter optimized and trained by the genetic algorithm, an adaptive threshold is used to segment most of the weak point targets detected, and a fixed threshold is used to segment the point targets with relatively high signal noise to detect the target point.

The specific implementation details of each part are as follows:

1. Construct a genetic algorithm for optimal training

Morphological filter parameters are mainly composed of component values of structural elements, and its training and learning process is a multi-parameter optimization problem. For this reason, when they are mapped to the string structure data composed of genes in the genetic space, a multi-parameter encoding method is adopted. The present invention introduces a new interval discretization coding, assuming that each component of the individual B is all real numbers in the interval [x _min , x _max ]. At this time, first determine the length K of the constructed string, and then insert 2 ^K -2 points at equal intervals on [x _min , x _max ], and the distance between two adjacent points is

$\mathrm{<span\; class="notranslate">\; \&delta;</span>}<span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}$

Then 2 ^K points are selected on [x _min , x _max ], they are respectively x _min , x _min +δ, x _min +2δ,..., x _min +(2 ^K -1)δ=x _max , which 2 ^K points are represented by K-bit binary numbers respectively, namely

The interval discretization coding and the conventional coding are simulated respectively, and the simulation results are shown in Table 3 and Table 4. Table 3 is a comparison of the CPU time consumed when the algorithm converges to different fitness levels when the conventional coding genetic algorithm is used for training. Table 4 compares the individual fitness under a certain number of cycles when the present invention is trained with the conventional coding genetic algorithm. Table 3 reflects that the convergence speed of the interval discretization coding genetic algorithm in the present invention is faster than that of the conventional coding genetic algorithm, and Table 4 reflects that its optimization performance is better than that of the conventional coding genetic algorithm.

The genetic algorithm learning rules need to introduce relevant prior knowledge and statistical laws to be constrained, and provide an optimal standard (cost function) to guide the solution process. For the optimization training of filtering parameters, the most critical thing is that the nonlinear mapping output of the morphological filter should be as close as possible to the expected value of the training samples, that is, the optimal solution is required to be consistent with all examples and be an optimal description. At the same time, it is also necessary to take into account the maneuverability of the termination algorithm (shutdown criterion). Therefore, the present invention selects the square error cost function of the optimal solution target as the cost function of deviation correction and traction optimization search, which is more ideal, and is defined as follows:

$\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; L</span>}}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Here L is the number of training samples, d _k is the expected value of the k-th input signal corresponding to the output, and Y _k is defined as the maximum value of the output matrix of the Top-Hat morphological filter after the k-th training sample is input, as shown below:

When surplus operation

Y _k ＝max(F _k -(F _k ·B)(x))

closed remainder operation

Y _k ＝max((F _k ·B)(x)-F _k )

Then the fitness value function f _i =fitness(pop _i (t)) of individual pop _i (t) is defined as:

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; E.</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; /</span>\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; L</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

The chromosomes of the parents are crossed to produce offspring chromosomes under a certain probability, so that the offspring individuals inherit the basic characteristics of the parents. The present invention newly designs a primary-subordinate crossover operator. The idea is that in the early stage of the algorithm, the crossover operation is mainly aimed at high-weight components near the origin, so as to increase the generation speed and probability of high-quality solutions and accelerate the elimination of low-quality solutions, thereby Make the search domain turn to the solution space with optimization potential as early as possible; in the later stage of the algorithm, the crossover operation focuses on the low-weight components of individual edges to protect the optimized high-weight components, and gradually optimize the secondary components. For this reason, the structural elements of size n×n can be regarded as a series of rectangles centered on the origin, and the weight of a certain component is measured by the side length of the rectangle where it is located. The side length of the i×i rectangle in the structural element is defined as i.

Given two structural element individuals pop ₁ (t) and pop ₂ (t) on the vector set Q. The primary and secondary crossover probability is defined as follows:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&alpha;\; L</span>}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; ]</span>& \stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\\ {\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span>\mathrm{<span\; class="notranslate">\; \&beta;L</span>}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}<span\; class="notranslate">\; ]</span>& \stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; \&Greater\; Equal;</span>\mathrm{<span\; class="notranslate">\; \&eta;</span>}\end{array}\right.$

In the formula, η is the fitness constant, L is the side length of the rectangle where the individual component is located, K ₁ and K ₂ are constants, α, β>0 are weight constants, is the average fitness value of the group, defined as follows

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; f</span>}}<span\; class="notranslate">\; =</span>\sqrt{{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; +</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}$

When the crossover operation is performed each time, N/2 samples are randomly selected as the population from the population with the number of N by the roulette selection method with the probability shown in the genetic algorithm STEP3. The present invention adopts a new ergodic intersection rule as follows:

{crosspop _i , crosspop _(N/2)+i }=cop(newpop _i , newpop _s )1≤i≤N/2 Among them, cop() is the cross function, newpop _i is the i-th individual in the population, newpop _s is the individual selected from the population by the roulette selection method with the probability shown in the genetic algorithm STEP3 at each crossover operation, and crosspop _i and crosspop _N/2+i are newly generated individuals after each crossover operation. In this way, after each round of crossover operation, the number of chromosomes in the new population generated is the same as the number of chromosomes in the original population.

The mutation operator realizes the optimization and improvement of the population, repairs and supplements some genetic genes that may be lost during the crossover process, and restores the lost diversity of the population to avoid falling into local optimum.

The present invention adopts a new primary and secondary variation based on the optimization principle of priority first, weight priority. In the early stage of the algorithm, the variation operation is mainly aimed at the high weight components near the individual origin, so as to reduce the unnecessary impact on the secondary components at this stage. Blind search; in the late stage of the algorithm, the mutation operation is mainly aimed at the low-weight components of the individual edge, so as to maintain the optimized main components and gradually optimize the secondary components.

Define the primary-minor variation probability as follows:

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; L</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\frac{{\mathrm{<span\; class="notranslate">\; L</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ]</span>& \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&le;</span>\frac{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\\ {\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span>\frac{{\mathrm{<span\; class="notranslate">\; L</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{\mathrm{<span\; class="notranslate">\; 2</span>}\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; ]</span>\mathrm{<span\; class="notranslate">\; exp</span>}<span\; class="notranslate">\; [</span><span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&tau;</span>}}<span\; class="notranslate">\; ]</span>& \mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; ></span>\frac{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}\end{array}\right.$

In the formula, P ₁ and P ₂ are amplitude constants, σ is a constant to adjust the weight of individual parameters, τ is a time constant to adjust the decay speed of P _m with time, and T is the maximum algebra set by the genetic algorithm.

2. Segmentation based on adaptive threshold

According to the idea of using adaptive threshold to segment most of the detected weak and small point targets and using a fixed threshold to segment point targets with relatively high signal noise, the present invention adopts the definition of single frame detection probability, false alarm probability and signal-to-noise ratio The adaptive threshold is as follows:

$\mathrm{<span\; class="notranslate">\; \&nu;</span>}<span\; class="notranslate">\; =</span>\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; SNR</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&sigma;\&Phi;</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; )</span>& \mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; t</span>}\\ \mathrm{<span\; class="notranslate">\; T</span>}& \mathrm{<span\; class="notranslate">\; \&sigma;</span>}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; t</span>}\end{array}\right.$

Among them, p _d is the detection probability of a single frame, SNR is the signal-to-noise ratio, v is the detection threshold, u is the mean value of the noise after the background cancellation of an n×n image unit, and σ ² is the mean square error of the noise. The calculation method of u and σ ² is as follows

$\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; g</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; g\&Theta;B</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CirclePlus;</span>\mathrm{<span\; class="notranslate">\; B</span>}$

$\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 9</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 9</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 3</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

g is the gray level of the original image, and f is the gray level after the morphological filtering operation.

Using the present invention to filter a series of infrared weak point target images, the filtering results are shown in FIG. 2 . And compared with different structural element filters and different morphological operator filters, the results are shown in Table 1 and Table 2. Table 1 compares the single-frame filtering processing of 100 images with different signal-to-noise ratios by the present invention and other two filters adopting different structural elements. Among them, the way (1) is to use the Top-Hat morphological filter with fixed structural elements for filtering processing, the way (2) is to use the neural network to optimize the training of the Top-Hat morphological filter for structural elements to perform filtering processing, and the way ( 3) To perform filter processing with the present invention. Table 2 shows the comparison between the present invention and the single-frame filtering process using different morphological operator filters under different signal-to-noise ratios. It can be seen that the method of the present invention can remarkably improve the filtering performance of weak and small target images compared with filters using fixed structural elements, neural network optimization training structural elements and other morphological operators. Compared with the morphological filter that uses the neural network to optimize the training structure elements, the present invention can also detect some weak point targets that it cannot detect, as shown in FIG. 3 .

Compared with the existing infrared weak and small target detection technology, the present invention can effectively suppress various noise interference, accurately detect point targets with low signal-to-noise ratio, and achieve engineering practical effect. At the same time, it can be seen from the whole implementation steps that the method of the present invention is easy to implement, thus providing a technical implementation method for the engineering of infrared weak point target detection.

Table 1 Comparison table of detection target probabilities of filters with different structural elements

SNR The detection probability of point target under mode (1) The detection probability of point target under mode (2) The detection probability of point target under mode (3) 1.5 90% 95% 97%

2 94% 96% 98% 5 100% 100% 100%

Table 2 Comparison table of detection target probability of different morphological operator filters

SNR Detection probability of point target under dilation operation Detection probability of point target under open operation Detection probability of point target under Top-Hat calculation 2 81% 92% 98% 3 84% 95% 100%

Table 3 Convergence schedule of genetic algorithm with different coding methods

adaptability CPU time used for regular encoding CPU time occupied during interval discretization encoding 0.3355 61.7720 23.2340 0.3550 114.8130 39.0310

Table 4 Comparison table of genetic algorithm optimization training performance with different coding methods

Cycles individual fitness Individual fitness in interval discretization coding 10 0.3109 0.3637 20 0.3264 0.3659

Claims (3)

1, a kind of morphological filter automatic target detection method, it is characterized in that, genetic algorithm is combined with Top-Hat morphological filter organically, at first collection is used for optimizing the training sample of training structural element, these samples should be as far as possible Contains various point targets and backgrounds, and constructs a genetic algorithm for optimal training. The genetic algorithm uses a new interval discretization coding and adaptive primary and secondary crossover and mutation operators, and uses the collected samples to optimize the genetic algorithm. Train the value of the structural elements, construct a morphological filter based on the Top-Hat operator on the basis of these optimized structural elements, filter the infrared target image, and finally perform self-adaptive based on most of the detected weak point targets Threshold segmentation, for point targets with relatively high signal noise, use a fixed threshold to segment and detect target points; The new interval discretization coding is specifically as follows: assuming that each component of the individual B takes the value of all real numbers in the interval [x _min , x _max ], first determine the length K of the constructed string, and then in [x _min , x max ], x _max ] to insert 2 ^K -2 points at equal intervals, and the distance between every two adjacent points is $\mathrm{\&delta;}=\frac{x_{\max }-x_{\min }}{2^K-1},$ Then 2 ^K points are selected on [x _min , x _max ], they are respectively x _min , x _min +δ, x _min +2δ,..., x _min +(2 ^K -1)δ=x _max , which 2 ^K points are represented by K-bit binary numbers respectively, namely

2. the morphological filter automatic target detection method according to claim 1, is characterized in that, described structure is used for optimizing the genetic algorithm of training, specifically: Including the steps of encoding, individual fitness calculation, crossover and mutation, among them, the encoding method adopts a new interval discretization encoding, which is equivalent to discretizing the real number range of the value of the individual component, and then calculating the value of each chromosome in the population The fitness function value of $p_1=\frac{\exp [f_1/T_k]}{\underset{J=1}{\overset{N}{\mathrm{\&Sigma;}}}\exp [f_J/T_k]}$ , i=1, 2,..., N use the roulette selection method to randomly select some chromosomes from the population to form a population, and finally generate a new population through adaptive primary and secondary crossover and mutation operations, where: f ₁ is the population The fitness value function of each chromosome in ; {T _k } is the annealing temperature that gradually tends to 0, and k=1, 2, . . . 3. The morphological filter automatic target detection method according to claim 1, wherein the segmentation based on the adaptive threshold is specifically: According to the noise mean value, noise mean square error, single frame detection probability and signal-to-noise ratio of an n×n image unit after background cancellation, the adaptive threshold value is calculated, and then the mean square error of each n×n image unit is given A threshold, most of the weak point targets whose noise mean square error is lower than this threshold are segmented by adaptive threshold, and point targets higher than this threshold are segmented by fixed threshold.

Images