CN106709497A - PCNN-based infrared motion weak target detection method - Google Patents

Abstract

The invention discloses a PCNN-based infrared motion weak target detection method. Filtering pretreatment is carried out on an infrared image by using a visual side suppression method, wherein compared with the conventional pretreatment method, the visual side suppression method has high background suppression and contrast enhancement capabilities and thus the background suppression and contrast enhancement effects are good; target segmentation is carried out on the image by using a PCNN method and preliminary separation is carried out on possible targets, background clutters and noise to determine a candidate target, wherein target point segmentation can be carried out accurately and a few of false alarm target points exist based on pulse coupling, nonlinear multiplication modulation, and variable threshold and other characteristics of the PCNN; and motion characteristics of the candidate target including a true target and a high-frequency noise point are analyzed by using an improved neighboring region judgment method, the size of a judged neighboring region is determined adaptively based on the motion speed of the target, and a true target is extracted by combining a multi-frame image flow analysis and high-frequency noises are removed.

Description

A kind of infrared moving detection method of small target based on PCNN

Technical field

The present invention relates to a kind of detection method of Moving Small Targeties in infrared image sequence, and in particular to one kind is based on arteries and veins The infrared moving detection method of small target of coupled neural network (PCNN, Pulse Coupled Neural Network) is rushed, Belong to technical field of computer vision.

Background technology

At present, infrared motion target detection method mainly faces two problems：(1) background is complicated.Background clutter is often It is motion, uneven and high brightness；(2) target is small and weak.Under normal circumstances, target sizes only have several to more than ten pixel, And brightness is relatively low.At present, it is conventional infrared based on single-frame images for the infrared image with strong clutter background and Weak target Object detection method is (as being based on the airspace filter method of high-pass filtering template, based on wavelet transformation and the morphologic filtering of mathematics Method etc.) antijamming capability it is poor, thereby increases and it is possible to produce false target, it is difficult to obtain preferable detection probability and false alarm rate.Together When, although the existing infrared target detection method based on sequence image can solve the above problems to a certain extent, deposit In certain limitation.For example：Frame difference method is easily influenceed by the interference of complex background and noise；The method requirement back of the body based on conversion Scape is uniform, and only target motion；The three-dimensional known priori of matched filtering method requirement, priori is inaccurate will to cause inspection Survey hydraulic performance decline；Hough transform method operand is larger, it is difficult to real-time implementation.Meanwhile, under Low SNR, its target Track is easily flooded by noise.Above mentioned problem causes that the infrared moving Dim targets detection under strong noise jamming turns into urgently to be resolved hurrily Problem, and increasingly paid attention to by researcher.

Neighborhood logic method is a kind of simple structure, function admirable and the infrared motion target detection method being easily achieved, Solve the problems, such as that the infrared moving Dim targets detection under strong noise jamming has advantage.Wherein, Target Segmentation is neighborhood logic method Key link, its accuracy directly influences object detection and recognition probability.Traditional image partition method is (such as between maximum kind Variance method, maximum entropy method (MEM) etc.) mainly to be split according to the appropriate threshold value of the gray difference of target and background selection, it is to parameter Selection have very strong dependence, robustness is poor, it is difficult to realize the segmentation of complex background, Weak target.On the other hand, as To the processing mode of information on the brain visual cortexs of animal such as third generation neutral net, PCNN modelings cat, monkey, in figure As segmentation aspect has various unique characteristics.For example：Such as pulse-couple characteristic, non-linear multiplication modulating characteristic, change threshold value spy Property and the neighborhood of neuron catch characteristic etc..There are some researches show above characteristic is conducive to intactly retaining the region of image Information, it is possible to resolve the segmentation problem of different infrared targets, and threshold value need not be adjusted with the change of image.Therefore, base is studied In PCNN and the infrared target detection method of neighborhood logic, for solving the infrared moving Dim targets detection under strong noise jamming Problem is significant.

The content of the invention

For detection in infrared image sequence with complex background Dim targets detection problem, the invention discloses a kind of based on the infrared of PCNN Moving Weak Small Target, it is intended to which the technical problem of solution is the accurate of raising infrared moving detection method of small target Property, robustness and antijamming capability.

The purpose of the present invention is achieved through the following technical solutions.

The invention discloses a kind of infrared moving detection method of small target based on PCNN.The method is first with vision Lateral inhibition method is filtered pretreatment to infrared image.Because vision lateral inhibition method has relative to conventional preprocess method Stronger suppression background, enhancing contrast ability, so as to can obtain more preferable background suppression, contrast enhancing effect；Then, profit Target Segmentation is carried out to image with PCNN methods, it would be possible to target and background clutter and noise initial gross separation, determine candidate's mesh Mark.The characteristics such as pulse-couple characteristic, non-linear multiplication modulating characteristic, change threshold property that PCNN has can split impact point Accurately, false-alarm targets point is less；Finally, using improved neighborhood logic method analysis candidate target (including real target and High-frequency noise point) kinetic characteristic, the adaptive Size of Neighborhood that should determine that judgement of movement velocity according to target, with reference to multiple image Flow point analysis extracts real target, rejects high-frequency noise.

The invention discloses a kind of infrared moving detection method of small target based on PCNN, comprise the following steps：

Step 1：Infrared image sequence is pre-processed.

It is right using wave filter (such as high-pass filtering, morphologic filtering, Butterworth filtering and filtering based on lateral inhibition) Image is filtered treatment, filters the background of smooth variation, extracts the high fdrequency components such as impact point and high-frequency noise.Due to vision Lateral inhibition method has relative to other wave filters preferably suppress background and enhancing contrast effect, preferably lateral inhibition method Infrared image is pre-processed.

First, the lateral inhibition Filtering Template that size is M × M is taken, wherein M values are bigger, and the inhibition to image is more obvious, But as M values become big, processing speed can be slack-off.Accordingly, it would be desirable to select suitable M values.

Then, using lateral inhibition template to original image in each pixel be filtered, filter process such as formula (1)：

I₁(x, y)=I (x, y)-I (x, y) * L (x, y) (1)

In formula, I (x, y) is the intensity profile of input picture, I₁(x, y) is by the output figure after lateral inhibition template convolution The intensity profile of picture, L (x, y) is lateral inhibition Filtering Template.

Step 2：Enter row threshold division using PCNN, determine candidate target.

Using PCNN methods will likely target and background clutter and noise carry out initial gross separation, determine candidate target. PCNN realizes that the process of Target Segmentation is as follows：

Step 2.1：Parameter in PCNN models is set.

Shown in the discrete type mathematical description such as formula (2) that single neuronal function is realized in PCNN models：

In formula, F_ijN () is the feed back input of (i, j) individual neuron, I_ijStimulus signal is an externally input, i.e. input figure The intensity profile of picture；L_ijN () is be of coupled connections input, Y_klThe output of neuron, W during for (n-1) secondary iteration_klIt is link weights Matrix；U_ijN () represents the internal activity of neuron, β is the bonding strength coefficient between cynapse；θ_ijN () is inside neurons Movable dynamic threshold, τ_θIt is dynamic threshold θ_ijTime attenuation constant；Y_ijN () is the time sequential pulse of PCNN outputs, work as neuron Internal activity U_ijN () is more than dynamic threshold θ_ijN when (), neuron excites generation pulse to export, and otherwise neuron is not excited, Pulse is not produced to export.

, it is necessary to arrange parameter during using the model, including：Link weight matrix W_kl, bonding strength factor beta and time decays Constant, τ_θ.Typically can based on experience value determine above parameter.

Step 2.2：Operation in PCNN networks is input an image into, candidate target is extracted.

By input picture I_ijStimulate as outside input, input picture is input in PCNN networks and is run.PCNN models Determine whether to light a fire the pixel according to the intensity profile of each pixel itself and its peripheral region, so that after being split Bianry image, and extract candidate target.

Step 3：Real target is extracted using improved neighborhood logic method.

The continuity and relevance moved using interframe target, by candidate target (including real target and high frequency Noise spot) Kinematic, select suitable neighborhood logic condition, extract real target with reference to Image flow analysis. Moving object detection step based on neighborhood logic method is as follows：

Step 3.1：Position according to candidate target determines the average movement velocity v of target.

First, component mark is attached to each two field picture in the bianry image sequence after segmentation.Wherein, adjoining side Formula (i.e. 4 adjoinings and 8 adjoinings) influences whether to connect the quantity of component, and 8 adjoinings can expand the model of connected region relative to 4 adjoinings Enclose, while reducing the quantity of connection component；It is the center for determining candidate target, next calculates in each two field picture and own Connect the barycenter of component and preserve；Then, the maximum abscissa of barycenter in each two field picture is taken respectively with former frame barycenter most Big abscissa subtracts each other, and obtains the matrix of each frame target speed, and the wherein unit of movement velocity is pixel/frame；Finally, it is right Movement velocity matrix takes median, used as the average movement velocity v of Weak target in the image sequence.

Step 3.2：According to candidate target barycenter whether in corresponding neighborhood, real target or high-frequency noise are judged.

The average movement velocity v calculated using step 3.1 determines to find the Size of Neighborhood S of candidate target, its calculating side Method such as formula (3)：

S=2v+1 (3)

In order to using the continuity and relevance of the motion of interframe target, carry out one to the M two field pictures in infrared sequence every time Secondary judgement, M values too senior general's influence arithmetic speed and efficiency are too small to decline accuracy of detection.The span of M is generally 7- 9。

In kth (k >=M) frame in, candidate target point P (x, y) is selected, judge S × S neighborhoods of (x, y) point in k-l frames Inside whether there is candidate target point.If so, then counter n+1 and continuing to judge k -2 frames；Otherwise, neighborhood is expanded as into (x, y) point (S+1) × (S+1) neighborhood, continues the judgement of the frame of kth -2.Interference in view of environment may make target in the image planes of a certain frame Be up to t frames target temporary extinction is allowed in temporary extinction, therefore every M frames, i.e.,：This step operation limited number of times is t times (t≤2)； If in M frame in counters n >=(M-t), judging the barycenter that it is real target.Otherwise, then noise spot is regarded it as.

Step 3.3：According to the target centroid for finding, the region where determining real target.

After determining the barycenter of real target, then the connection where finding the barycenter in the mark image in step 3.1 Region where region, i.e. real target, so that by real Objective extraction out.

Beneficial effect：

1st, strong robustness.A kind of infrared moving detection method of small target based on PCNN disclosed by the invention, utilizes PCNN methods can realize the Threshold segmentation under conditions of noise jamming, and need not adjust segmentation threshold；Meanwhile, in neighborhood logic side The Size of Neighborhood adjudicated according to target speed Automatic adjusument in method, so that suitable for the moving target inspection of friction speed Survey, effectively increase the robustness of algorithm.

2nd, strong antijamming capability.A kind of infrared moving detection method of small target based on PCNN disclosed by the invention, depending on Feel that lateral inhibition method has preferably suppression background and enhancing contrast effect relative to conventional preprocess method, can effectively press down Make complicated background clutter；Whether candidate target barycenter is judged in corresponding neighborhood according to the continuity that candidate target is moved, So as to judge target or high-frequency noise, the interference of the factors such as high-frequency noise, cloud and mist can be effectively antagonized.

3rd, false alarm rate is low.A kind of infrared moving detection method of small target based on PCNN disclosed by the invention, it is improved Whether neighborhood logic method judges candidate target barycenter in corresponding neighborhood, so as to sentence according to the continuity that candidate target is moved Break and target or high-frequency noise, so that the method has rejects high-frequency random noises ability, can effectively extract Weak target, So as to reduce false alarm rate.

Brief description of the drawings

Fig. 1 is the flow of moving target detecting method of the invention；

Fig. 2 is the flow that neighborhood logic of the invention extracts real target；

Fig. 3 is the detection knot of threshold segmentation method of the invention and regulatory thresholds dividing method (OTSU methods and maximum entropy method (MEM)) Fruit mass ratio compared with；

Fig. 4 is the frame original image of the 16th, 17,26 and 27 and its motion detection result in list entries.

Specific embodiment

In order to better illustrate objects and advantages of the present invention, present invention is done into one with example below in conjunction with the accompanying drawings Step explanation.

Embodiment 1：

A kind of infrared moving detection method of small target based on PCNN, its overall process such as accompanying drawing disclosed in the present embodiment Shown in 1, following steps are specifically included：

Step 1：Infrared image sequence is pre-processed.

Treatment is filtered to image using based on the theoretical wave filter of lateral inhibition, the background of smooth variation is filtered, extracted Go out the high fdrequency components such as impact point and high-frequency noise.The lateral inhibition Filtering Template that size is 5 × 5, wherein lateral inhibition filtering are taken first Shown in template such as formula (4)：

Using the lateral inhibition template shown in formula (4) to original image in each pixel be filtered, filter process Such as formula (5)：

I₁(x, y)=I (x, y)-I (x, y) * L (x, y) (5)

In formula, I (x, y) is the intensity profile of input picture, I₁(x, y) is by the output figure after lateral inhibition template convolution The intensity profile of picture, L (x, y) is lateral inhibition Filtering Template.

Step 2：Enter row threshold division using PCNN, determine candidate target.

Using PCNN method will likely target and background clutter and noise initial gross separation, determine candidate target.Utilize The step of PCNN realizes Target Segmentation is as follows：

Step 2.1：The parameter of PCNN models is set.

In PCNN models, shown in the discrete type mathematical description such as formula (6) that single neuronal function is realized：

In formula, F_ijN () is the feed back input of (i, j) individual neuron, I_ijStimulus signal is an externally input, i.e. input figure The intensity profile of picture；L_ijN () is be of coupled connections input, Y_klThe output of neuron, W during for (n-1) secondary iteration_klIt is link weights Matrix；U_ijN () represents the internal activity of neuron, β is the bonding strength coefficient between cynapse；θ_ijN () is inside neurons Movable dynamic threshold, τ_θIt is dynamic threshold θ_ijTime attenuation constant；Y_ijN () is the time sequential pulse of PCNN outputs, work as neuron Internal activity U_ijN () is more than dynamic threshold θ_ijN when (), neuron excites generation pulse to export, and otherwise neuron is not excited, Pulse is not produced to export.

Arrange parameter is needed when using the model, including：Link weight matrix W_kl=[0.07 0.1 0.07；0.1 0 0.1；0.07 0.1 0.07], bonding strength factor beta=0.4 and time attenuation constant τ_θ=0.3.

Step 2.2：Operation in PCNN networks is input an image into, candidate target is extracted.

By input picture I_ijStimulate as outside input, input an image into operation, PCNN model foundations in PCNN networks The intensity profile of each pixel itself and its peripheral region determines whether to light a fire the pixel, so that the two-value after being split Image, and extract candidate target.

The specific embodiment of step 2.2 is followed the steps below, and detailed process is illustrated by taking the 1st iteration as an example：

2.2a) receive input domain.

Using the image after step 1 is processed as the primary input (F) of PCNN, the anti-noise of PCNN can be so significantly improved Ability, such as formula (7)：

F_ij(1)=I_ij1 (7)

In formula, I_ij1Represent that, by the image after step 1 treatment, (1) represents the 1st iteration, F_ij(1) it is (i, j) individual god Through the primary input of unit.

Weighting by the use of output signal and link weights W is input into (L) as the link of PCNN, such as formula (8)：

In formula, weight matrix W is linked_kl=[0.07 0.1 0.07；0.1 0 0.1；0.07 0.1 0.07], using its energy The 8- neighborhood territory pixels value of environmental stimuli is chained up by weight size, Y_klThe output of neuron during for the 0th iteration, therefore just Initial value is 0；

2.2b) adjust link field.

In modulation link field, primary input and link input to being received into coming carry out global modulation coupling, such as formula (9) institute Show：

In formula, U_ij(1) internal activity of neuron is represented, β is that the bonding strength coefficient value between cynapse is 0.4, Its value is bigger, represents that the neuron in 8- neighborhoods is bigger on the influence of its central nervous unit.

2.2c) pulses generation domain.

In pulses generation domain, output signal is relatively determined by internal activity and dynamic threshold, work as inside neurons When active entry is more than dynamic threshold, neuron excites generation pulse to export, and otherwise neuron is not excited, and does not produce pulse to export, Shown in its dynamic threshold and output pulse such as formula (10) and (11)：

θ_ij(1)=exp (- τ_θ)θ_ij(0) (10)

In formula, θ_ij(1) it is inside neurons activity dynamic threshold, τ_θIt is dynamic threshold θ_ijTime attenuation constant value be 0.3；Y_ij(1) it is the time sequential pulse of PCNN outputs.

Step 3：Real target is extracted using improved neighborhood logic method.

The continuity and relevance moved using interframe target, by candidate target (including real target and high frequency Noise spot) Kinematic, select suitable neighborhood logic condition, extract real target with reference to Image flow analysis, Its detailed process is as shown in Figure 2.

Moving object detection step based on neighborhood logic method is as follows：

Step 3.1：Position according to candidate target determines the average movement velocity v of target.

First, each frame of the bianry image sequence after segmentation is labeled the operation of connection component according to 8 adjoinings；For Determine the center of candidate target, next calculate the barycenter of all connection components in each two field picture and preserve；Then, take The maximum abscissa of barycenter subtracts each other with the maximum abscissa of former frame barycenter respectively in each two field picture, obtains rough each frame The unit of the matrix of target speed, wherein movement velocity is pixel/frame；Finally, median is taken to movement velocity matrix, is made It is the average movement velocity v of Weak target in the image sequence.

By taking the image of continuous 30 frame as an example, by calculate the maximum abscissa of each frame candidate target barycenter respectively with it is previous The matrix that the difference of the maximum abscissa of frame barycenter obtains each frame target speed is：[0 0 4 0 -6 9 3 -7 4 -8 0 1-6 8-5-3 370271 7-2 21 7-4 3-7], then it is reduction error, it is to avoid accidental sexual factor, Average movement velocity v=1 of the median as Weak target in the image sequence is taken to movement velocity matrix.

Step 3.2：According to candidate target barycenter whether in corresponding neighborhood, real target or high-frequency noise are judged.

It may be real target by the candidate target being partitioned into, it is also possible to high frequency very noisy point.Target is in interframe Motion has continuity and relevance；And the motion of noise is random, interframe does not have relevance.Thus, it is possible to pass through right The analysis of the position relationship (movable information) of continuous interframe candidate point target, extracts real target, i.e.,：If candidate's point target Occur in 8 neighborhoods of previous frame image same position, then retain；Otherwise it is noise, the pixel zero setting.In view of the interference of environment Target temporary extinction in the image planes of a certain frame may be made.If therefore it is presumed that certain candidate target has at least 6 in continuous 8 frame Frame occurs, then be judged as real target.

First, the average movement velocity v for being calculated using step 3.1 determines to find the Size of Neighborhood S of candidate target, its meter Calculation method such as formula (12)：

S=2v+1 (12)

Still by taking the image of continuous 30 frame as an example, the Size of Neighborhood of candidate target is calculated as S=2 × 1+1 according to formula (12) =5, it is the continuity and relevance moved using interframe target, the M two field pictures in infrared sequence are once judged every time, M=8 is taken in the present invention.

Next, by taking the 16th frame in continuous 30 frame as an example, candidate target point P (x, y) is selected in the 16th frame, judge Whether there is candidate target point to occur in 5 × 5 neighborhoods of (x, y) point in 15th frame, if counter Num=Num+1 if having (at the beginning of Num 0) initial value is and continues to judge 14 frames；Otherwise, neighborhood is expanded as 6 × 6 neighborhoods of (x, y) point, continues the judgement of 13 frames, this step Operation limited number of times may make target temporary extinction in the image planes of a certain frame for 2 times, the i.e. interference in view of environment.Therefore, Be up to 2 frame target temporary extinctions are allowed in every 8 frame；If in 8 frame in counter Num >=6, being judged as the matter of real target The heart, is otherwise considered as noise spot.

Step 3.3：Target centroid according to finding determine real target where region.

After determining the barycenter of real target, then the company where can finding the barycenter in the mark image in step 3.1 Region where logical region, i.e. real target, so that by real Objective extraction out.

Fig. 3 is threshold segmentation method of the invention and regulatory thresholds dividing method (OTSU methods and maximum variance between clusters) Testing result mass ratio compared with.Wherein, the image in Fig. 3 (a)-(d) is from left to right followed successively by original target image, OTSU methods, most The testing result of big Ostu method and the inventive method.It can be seen that can preferably be pressed down using the inventive method The interference of background clutter processed, extracts Weak target and reduces noise as far as possible, and its effect is better than other four kinds of methods.

Fig. 4 is the frame original image of the 16th, 17,26 and 27 and its motion detection result of list entries.Can from figure Go out, the neighborhood logic method after improvement can make accurately judgement to real goal and high-frequency noise, so as to effectively reduce false-alarm Rate.

Above-described specific descriptions, purpose, technical scheme and beneficial effect to inventing are further elaborated, Should be understood that and the foregoing is only specific embodiment of the invention, the protection domain being not intended to limit the present invention, All any modification, equivalent substitution and improvements within the spirit and principles in the present invention, done etc., should be included in of the invention Within protection domain.

Claims (5)

1. a kind of infrared moving detection method of small target based on PCNN, it is characterised in that：

Comprise the following steps：

Step 1：Infrared image sequence is pre-processed.

Using wave filter (such as high-pass filtering, morphologic filtering, Butterworth are filtered and based on the theoretical filtering of lateral inhibition) to figure As being filtered treatment, the background of smooth variation is filtered, extract the high fdrequency components such as impact point and high-frequency noise.

Step 2：Enter row threshold division using PCNN, determine candidate target.

Using PCNN method will likely target and background clutter and noise initial gross separation, determine candidate target.PCNN is realized The process of Target Segmentation comprises the following steps：

Step 2.1：Parameter in PCNN models is set.

Shown in the discrete type mathematical description such as formula (1) that single neuronal function is realized in PCNN models：

$\left\{\begin{array}{c}{F}_{ij}\left(n\right)=I_{ij}\\ L_{ij}\left(n\right)=\underset{k,l}{\&Sigma;}{W}_{kl}{Y}_{kl}(n-1)\\ U_{ij}\left(n\right)=F_{ij}\left(n\right)(1+{\mathrm{\&beta;L}}_{ij}(n\left)\right)\\ {\mathrm{\&theta;}}_{ij}\left(n\right)=\exp (-{\mathrm{\&tau;}}_{\mathrm{\&theta;}}){\mathrm{\&theta;}}_{ij}(n-1)\\ Y_{ij}\left(n\right)=\left\{\begin{array}{cc}1& U_{ij}\left(n\right)\&GreaterEqual;{\mathrm{\&theta;}}_{ij}\left(n\right)\\ 0& U_{ij}\left(n\right)<{\mathrm{\&theta;}}_{ij}\left(n\right)\end{array}\right.\end{array}\right.---\left(1\right)$

In formula, F_ijN () is the feed back input of (i, j) individual neuron, I_ijIt is an externally input stimulus signal, i.e. input picture Intensity profile；L_ijN () is be of coupled connections input, Y_klThe output of neuron, W during for (n-1) secondary iteration_klIt is link weight matrix； U_ijN () represents the internal activity of neuron, β is the bonding strength coefficient between cynapse；θ_ijN () is inside neurons activity Dynamic threshold, τ_θIt is dynamic threshold θ_ijTime attenuation constant；Y_ijN () is the time sequential pulse of PCNN outputs, work as inside neurons Active entry U_ijN () is more than dynamic threshold θ_ijN when (), neuron excites generation pulse to export, and otherwise neuron is not excited, and is not produced Raw pulse output.

Arrange parameter is needed when using the model, including：Link weight matrix W_kl, bonding strength factor beta and time decay are normal Number τ_θ, typically can based on experience value determine above parameter.

Step 2.2：Operation in PCNN networks is input an image into, candidate target is extracted.

By input picture I_ijStimulate as outside input, be then enter into being run in PCNN networks, PCNN models are according to every The intensity profile of individual pixel itself and its peripheral region determines whether to light a fire the pixel, so that the binary map after being split Picture, and extract candidate target.

Step 3：Real target is extracted using improved neighborhood logic method.

The continuity and relevance moved using interframe target, by candidate target (including real target and high-frequency noise Point) Kinematic, select suitable neighborhood logic condition, extract real target with reference to Image flow analysis.

2. a kind of infrared moving detection method of small target based on PCNN as claimed in claim 1, it is characterised in that：

The moving object detection step based on improved neighborhood logic method described in step 3 is as follows：

Step 3.1：Position according to candidate target determines the average movement velocity v of target.

Each frame of the bianry image sequence after segmentation is labeled the operation of connection component, wherein abutment (i.e. 4 first Adjacent and 8 adjoinings) influence whether to connect the quantity of component, 8 adjoinings can expand the scope of connected region relative to 4 adjoinings, while Reduce the quantity of connection component；To determine the center of candidate target, all connections point in each two field picture are next calculated The barycenter of amount is simultaneously preserved；Then take the maximum abscissa of barycenter in each two field picture respectively with the maximum abscissa of former frame barycenter Subtract each other, obtain the matrix of rough each frame target speed, the unit of wherein movement velocity is：Pixel/frame；Finally to fortune Dynamic rate matrices take average movement velocity v of the median as Weak target in the image sequence.

Step 3.2：Whether real target or high-frequency noise are judged in corresponding neighborhood according to candidate target barycenter.

The average movement velocity v calculated using step 3.1 determines to find the Size of Neighborhood S of candidate target, and its computational methods is such as Formula (2)：

S=2v+1 (2)

It is the continuity and relevance moved using interframe target, the M two field pictures in infrared sequence are once judged every time, M values too conference influence arithmetic speed and efficiency, too small and can reduce the degree of accuracy, general value is 7-9.

Select candidate target point P (x, y) in kth (k >=M) frame, judge (x, y) point in k-l frames S × S neighborhoods it is interior whether There is candidate target point to occur, counter Num+1 and continue to judge k -2 frames if having；Otherwise neighborhood is expanded, that is, expand as (x, Y) (S+1) × (S+1) neighborhoods of point, continue the judgement of k -2 frames, and this step operation limited number of times is t times (t≤2), i.e., in view of ring The interference in border may make to allow be up to t frames target temporarily to disappear in target temporary extinction in the image planes of a certain frame, therefore every M frames Lose；If in M frame in counters Num >=(M-t), being judged as the barycenter of real target, otherwise it is considered as noise spot.

Step 3.3：Target centroid according to finding determine real target where region.

After determining the barycenter of real target, then the connected region where can finding the barycenter in the mark image in step 3.1 Region where domain, i.e. real target, so that by real Objective extraction out.

3. a kind of infrared moving detection method of small target based on PCNN as claimed in claim 1, it is characterised in that：

Wave filter described in step 1 image is filtered treatment can for high-pass filtering, morphologic filtering, Butterworth filtering and Based on theoretical filtering of lateral inhibition etc., due to vision lateral inhibition method have relative to other wave filters preferably suppress background and Enhancing contrast effect, preferably lateral inhibition method is filtered pretreatment to infrared image.

The lateral inhibition Filtering Template that size is M × M is taken first, and each pixel during then lateral inhibition template is to original image is filtered Ripple, filter process such as formula (3)：

I₁(x, y)=I (x, y)-I (x, y) * L (x, y) (3)

In formula, I (x, y) is the intensity profile of input picture, I₁(x, y) is by the output image after lateral inhibition template convolution Intensity profile, L (x, y) is lateral inhibition Filtering Template.

4. a kind of infrared moving detection method of small target based on PCNN as claimed in claim 3, it is characterised in that：

Lateral inhibition Filtering Template size described in step 1 is M × M, and wherein M values are bigger, and the inhibition to image is more obvious, but It is that processing speed can be slack-off, it is therefore desirable to selects suitable M values as M values become big.It is 5, wherein lateral inhibition that the present invention takes M values Shown in Filtering Template such as formula (4)：

$L=\left[\begin{array}{ccccc}0.025& 0.025& 0.025& 0.025& 0.025\\ 0.025& 0.075& 0.075& 0.075& 0.025\\ 0.025& 0.075& 0& 0.075& 0.025\\ 0.025& 0.075& 0.075& 0.075& 0.025\\ 0.025& 0.025& 0.025& 0.025& 0.025\end{array}\right]---\left(4\right)$

5. a kind of infrared moving detection method of small target based on PCNN as described in claim 1,2,3 or 4, its feature exists In：

Pretreatment is filtered to infrared image using vision lateral inhibition method, because vision lateral inhibition method is relative to others Preprocess method has preferably suppress background and enhancing contrast effect, suppresses so as to obtain background, the enhanced figure of contrast Picture；Target Segmentation is carried out to image using the method for PCNN, it would be possible to target and background clutter and noise initial gross separation, it is determined that Candidate target, many good characteristics such as pulse-couple characteristic, non-linear multiplication modulating characteristic, change threshold property etc. of PCNN are caused Impact point segmentation is accurate, false-alarm targets point is less；Using improved neighborhood logic method to candidate target (including real target With high-frequency noise point) Kinematic, the adaptive Size of Neighborhood that should determine that judgement of movement velocity according to target, with reference to many The analysis of two field picture flow point more stably extracts real target, weeds out high-frequency noise.