CN1216343C - Infrared target identification method based on unchanged rotary morphology neural net - Google Patents

Abstract

The present invention relates to an infrared target identification method based on an unchanged rotary morphology neural network. The mode of a traditional structure pattern library is improved. An incomplete mode which satisfies a topological structure is stored in the grading method in the meantime of storing a complete mode. The division of regional areas of the target is realized according to apriori information of the occurring position of a target which is provided by a target detecting stage to reduce the work load of the division which is carried out to a complete image. The nature that a Zernicke moment is rotary and unchanged is used for standardizing the target which is divided. After a target which is standardized is obtained, the target is combined with the information which is provided by the Zernicke moment. The shape information of the target is extracted according to the position of the target which is positioned by a nearest neighbour method in the mode library and the conversion principle of hit and miss in themorphology. An outputted result of the extraction serves as the input of a classifying network. At last, the correct classification of the target is realized. The present invention can realize the recognition when an object in the air is far from an imaging sensor, the shape information of the target is not complete, and a changing situation happens to the posture of the target.

Description

Infrared Target Recognition Method based on invariable rotary morphology neural network

Technical field:

The present invention relates to a kind of Infrared Target Recognition Method based on invariable rotary morphology neural network, can realize that aerial object is far away apart from imaging sensor, identification when the imperfect and attitude of target shape information changes situation, be a core technology of infrared detection and recognition system, infraed early warning system, big visual field targeted surveillance system, in all kinds of military, civilian systems, all can be widely used.

Background technology:

Infrared imagery technique is a kind of contactless measuring technology, and it can detect the invisible heat radiation that target sends easily and convert visual picture to.Information is obtained the related research that key area is infrared detection technique and method, and its critical role becomes increasingly conspicuous.Owing to advantages such as infrared imagery technique has good concealment, investigative range is wide, bearing accuracy is high, identification camouflage ability is strong, operating distance is far away and lightweight is small and exquisite, low consumption is reliable enjoy favor, can be widely used in fields such as security monitoring, national defense and military and industrial automation detection.

Along with the development of infrared eye technology, thermal imaging system adds one dimension or two-dimentional opto-mechanical scanner from past employing unit or polynary discrete detector, has developed into the gazing type imaging device without optical mechaical scanning.Based on the infrared thermal imaging detection system of staring focal plane arrays (FPA), no matter on temperature control and spatial resolution, still on frame frequency and spectral response, all be greatly improved.Since the focal plane stare thermal imaging system exclusive premium properties, become the new and high technology that research and develop energetically countries in the world.As one of key link of Intelligentized Information, infrared target imaging, detection and recognition technology are the bottleneck problem of puzzlement and restriction infrared imaging detection Practical Performance and technological difficulties and need to be resolved hurrily always, caused at present domestic and international expert's great attention, and carried out deep, extensive studies around this problem.

Infrared target detect with identifying in, need detect and lock target to be identified as soon as possible, but under the empty condition, when object far the time, realize that the purpose of discerning is faced with numerous technical barriers with imaging sensor.These technical barriers mainly contain:

1. target and imaging sensor are far away, and owing to reasons such as imaging sensor precision, target shape information is imperfect, and traditional target identification method can't be used;

2. the motion of imaging platform self vibration and target usually shows target attitude and changes, and has strengthened the difficulty of identification;

3. data volume is big, is difficult to real-time processing.

The researchist has proposed certain methods at Target Recognition both at home and abroad, as template matches, statistical model identification, sentence structure or tactic pattern identification and neural network recognition method, but but when target information is imperfect, or targeted attitude always shows the big problem of calculated amount when changing, and is difficult to handle in real time sequence of video images.

Summary of the invention:

The objective of the invention is to above-mentioned deficiency at prior art, a kind of recognition methods to the remote infrared target of sky-invariable rotary morphology neural net method is provided, this method can be carried out Digital Signal Analysis and Processing with imaging hardware system is supporting, improve the performance of system identification, satisfy performance requirements such as the recognition correct rate of real system and real-time.

For realizing such purpose, in the technical scheme of the present invention, improve structural model storehouse mode earlier, when depositing integrated pattern, will satisfy of the method storage of the imperfect pattern of topological structure by classification.Utilize this building method, each integrated pattern derives a series of imperfect pattern in the library, has constituted a group mode class.According to the prior imformation that the target detection stage can provide target the position to occur, realization target regional area is cut apart, to reduce the workload that entire image is cut apart.The motion of imaging platform self vibration and target usually shows target attitude and changes, and utilizes the character of Zelnick (Zernike) square invariable rotary, with the target specificationization that is partitioned into.This standardization processing has improved the comparability of each mode class in target to be identified and the known mode.After the target after obtaining standard, the information that provides in conjunction with the Zelnick square, according to the position of nearest neighbor method localizing objects in library, according to the conversion principle of " hitting miss " in the morphology target shape information is realized extracting, its output result finally realizes the correct classification of target as the input of sorter network.

Method of the present invention comprises following concrete steps::

1. creation mode storehouse.

Library is the basis of recognizer, and what the present invention is directed to is aerial target, then deposits various aerial targets in the library.The present invention makes full use of the topology information of pattern when the structural model storehouse, promptly not only deposit complete models in the library, but also deposit the non-complete models that satisfies the topological structure requirement, and arrange according to hierarchical pattern.A kind of integrated pattern in the library derives a group mode class like this.

2. according to the imaging characteristics of actual infrared image, utilize algorithm of target detection to realize target localization.

What the present invention is directed to is the infrared sequence Target Recognition, but when the target range imaging sensor is far, target shows as point target at imaging plane, at this moment because target does not still possess and can not discern needed information, can only utilize this moment algorithm of target detection to realize detecting and to target localization.

3. the target position information that provides of combining target detection-phase is realized the local segmentation of target.

Target is in motion process, and when dwindling gradually with imaging sensor distance, the shape information of target is complete gradually, and at this moment also to be increased gradually by point target be blob target to target, forms the appearance mark at last.The positional information of the target that the target detection stage provides can be used as the prior imformation that realizes that target is cut apart in local zonule.

4. utilize the rotational invariance of Zelnick square, with the target specificationization that is partitioned into.

For a width of cloth digital picture, the essence of calculating the Zelnick square is exactly the projection of image function at the Ze Er polynomial space, and the mould of result of calculation does not change with the image rotation.Zelnick square process of normalization is exactly the numerical value according to the different orders of Zelnick polynomial computation Zelnick square, and stores as a n dimensional vector n.

5. it is right to determine to be used in the feature extraction layer structural element of the morphology conversion of " hitting miss " according to nearest neighbor method.

The Zelnick square of the various mode class in the library is compared with target Zelnick square, utilize two norm distances of vector, realize the location of target in library in conjunction with nearest neighbor method, and it is right to be identified for the structural element of the morphology conversion of " hitting miss ".

6. normalized result is imported the feature extraction layer of invariable rotary neural network, utilize morphologic " hitting miss " conversion to realize the target shape feature extraction, and will calculate the eigenwert input category network of gained.The process sorter network calculates and finally obtains is the judgement of non-target.

The present invention at first improves traditional mode storehouse building method, and it is perfect to utilize topology information to carry out library, promptly not only deposits complete models in the library, but also deposits the non-complete models that satisfies the topological structure requirement, and arrange according to hierarchical pattern.Target detection result has realized the target localization function, utilizes this information to reach the purpose of target local segmentation, makes target detection and Target Recognition combine closely, and the process of Target Recognition is shifted to an earlier date as far as possible.Utilize the rotational invariance of Zelnick square that segmentation result is standardized, this normalized effect has simplified the difficulty in structural model storehouse, and a common position angle only needs a mode class, and needn't store the same pattern of a large amount of different rotary angles.According to the arest neighbors method, normalized result and library pattern relatively, it is right to determine that feature extraction layer is used for the structural element of the morphology conversion of " hitting miss ".As input, network structure then is to join the formula feedforward network entirely to sorter network with the feature extraction layer operation result.The present invention is under the condition that guarantees the algorithm identified performance, improved the real-time of algorithm, can realize that aerial object is far away apart from imaging sensor, identification when the imperfect and attitude of target shape information changes situation, can be widely used in all kinds of military, civilian systems, have vast market prospect and using value.

Description of drawings:

Fig. 1 is the invariable rotary morphology neural network structure figure to the remote infrared identification of sky.

As shown in Figure 1, for the true infrared image of a width of cloth, adopt the regional area of object detection method realization target to cut apart earlier.After the result of cutting apart utilized Zelnick square standardization, the input feature vector extract layer carried out the morphology conversion of " hitting miss ", and the k among the figure has represented the logarithm of structural element.The input of the output result of feature extraction layer as the feature extraction network.Finally drawing after sorter network through training calculates is the judgement of non-target.

Fig. 2 for when training sample and test samples correlativity than by force the time, be used for the true infrared image sample of invariable rotary morphology neural metwork training and supervising network performance.

Fig. 2 (a) figure belongs to target side to the flight training sample, and Fig. 2 (b), Fig. 2 (c) belong to the training sample of target forward flight, and the former shape is still imperfect, and latter's shape is complete substantially.

Fig. 3 is used for the true infrared image sample of invariable rotary morphology neural metwork training and supervising network performance for when training sample and test samples correlativity are more weak.

Fig. 3 (a) is the training sample, Fig. 3 (b), Fig. 3 (c) are the check sample, three kinds of sample standard deviations are from the same orientation angle, and difference is that test samples and the difference of the flight attitude in Fig. 3 (a) training sample among Fig. 3 (b) is bigger, a large amount of clutters occurred in the sample of Fig. 3 (c).

Embodiment:

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

The concrete implementation detail of each several part of Infrared Target Recognition Method that the present invention is based on invariable rotary morphology neural network is as follows:

1. creation mode storehouse.

To target imaging in the empty infrared image, because the phenomenon that target does not have perfect imaging can appear in the influence of factor such as imaging sensor imaging precision and target range be far away.But as long as the target of imaging has due topological structure, promptly inside, target area is smooth, does not exist very for a short time when empty, is to realize Target Recognition by utilizing these incomplete local messages.It is perfect that the present invention utilizes topology information to carry out library, promptly not only deposits complete models in the library, but also deposit the non-complete models that satisfies the topological structure requirement, and arrange according to hierarchical pattern.A kind of integrated pattern in the library derives a group mode class like this.The structural element that utilizes this library can be identified for very easily in " hitting miss " conversion of Shape Feature Extraction is right.

2. according to the imaging characteristics of actual infrared image, utilize algorithm of target detection to realize target localization.

When the target range imaging sensor was far, target showed as point target at imaging plane, utilized algorithm of target detection to realize target localization.Algorithm of target detection is an interframe correlativity when moving according to point target, and the feature that waits at random of background distributions homogeneity and noise profile realizes target travel flight path energy accumulation, improves the method for signal noise ratio (snr) of image, realizes Point Target Detection.

3. the target position information that provides of combining target detection-phase is realized the local segmentation of target.

Target is in motion process, and when dwindling gradually with imaging sensor distance, the shape information of target is complete gradually, and at this moment also to be increased gradually by point target be blob target to target, forms the appearance mark at last.The target position information that the combining target detection-phase provides, the local segmentation of realization target.Because target detection carries out the transition to the Target Recognition stage and do not have strict boundary, the present invention utilizes the target position information that the target detection stage obtains, and can realize in the zonule of position appears in target that target cuts apart.This method has been limited in this process in the less zone, has simplified the image segmentation complexity.The target location point that the target detection stage obtains is the barycenter of target to be identified not necessarily, but one fixes in the target area.Aerial target is in when motion and atmosphere friction and self be heater, makes target temperature higher with respect to surrounding environment.By the infrared thermal imaging principle, target presents high luminance area in infrared image, and surrounding environment belongs to the low-light level district.Utilize above-mentioned priori, can set a given threshold value with target detection point as seed points, the connection topological structure constraint that utilizes imageable target to have utilizes the seed growth method target can be split from background.The target approach cognitive phase, target need show information such as certain shape or texture.The present invention has utilized the prior imformation that the target detection stage provides, and algorithm design is from detecting this continuous process of identification, the target performance for by point target to little target, increasing gradually is that appearance is marked.When target increases to certain area, at this moment change cognitive phase over to by detection, this area can preestablish according to actual conditions.

4. utilize the rotational invariance of Zelnick (Zernike) square, with the target specificationization that is partitioned into.

Zelnick has been introduced one group of complex polynomial, and they have constituted unit circle (is x ²+ y ²=1) the complete orthogonal set in inside.If this group polynomial expression is by { V _Nm(x, y) } expression, this organizes polynomial form and is so:

V _Nm(x, y)=V _Nm(ρ, θ)=R _Nm(ρ) in exp (jm θ) following formula, n is a nonnegative integer, and m is not equal to 0 integer, and satisfies condition, n-|m| be even number and | m|≤n.ρ is that (x, vector length y), θ are vectorial ρ and x axle angle in the counterclockwise direction from the initial point to the pixel.

Definition is a formula R radially _Mn(ρ) as follows:

$R_{\mathrm{nm}}\left(\mathrm{\ρ}\right)=\underset{s=0}{\overset{(n-|m\left|\right)/2}{\mathrm{\Σ}}}{(-1)}^s\·\frac{(n-s)!}{s!(\frac{n+\left|m\right|}{2}-s)!(\frac{n-\left|m\right|}{2}-s)!}{\mathrm{\ρ}}^{n-2s}$

Following formula has R _N-m(ρ)=R _Nm(ρ).

This group polynomial expression is quadrature and satisfied,

${\∫\∫}_{x^2+y^2\≤1}{\left[V_{\mathrm{nm}}(x,y)\right]}^{*}{V}_{\mathrm{pq}}(x,y)\mathrm{dxdy}=\frac{\mathrm{\π}}{n+1}{\mathrm{\δ}}_{\mathrm{np}}{\mathrm{\δ}}_{\mathrm{mq}}$

Wherein, ${\mathrm{\δ}}_{\mathrm{mq}}=\left\{\begin{array}{cc}1& a=b\\ 0& \mathrm{otherwise}.\end{array}\right\}$

The Zelnick square is the projection of image function on these orthogonal basis functions.To the continuous images function carry out m recirculate outside unit circle disappear n rank Zelnick square be:

$A_{\mathrm{nm}}=\frac{n+1}{\mathrm{\π}}{\∫\∫}_{x^2+y^2\≤1}f(x,y)V_{\mathrm{nm}}^{*}(\mathrm{\ρ},\mathrm{\θ})\mathrm{dxdy}$

For a width of cloth digital picture, replace integration to obtain with summation,

$A_{\mathrm{nm}}=\frac{n+1}{\mathrm{\π}}\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}f(x,y)V_{\mathrm{nm}}^{*}{,x}^2+y^2\≤1$

Calculate the Zelnick square of a given image, as true origin, the coordinate Mapping of pixel is in the unit circle scope the center of image, and the pixel of those unit circle outsides does not participate in calculating.Convenience when the major reason of selection Zelnick square is its structure High Order Moment in numerous orthogonal moments be the more important thing is to be to be close to perfect invariable rotary feature.The mould of Zelnick square | A _Nm| have the rotational invariance characteristic, can be with | A _Nm| as the unchangeability feature of target.In the infrared image of reality, target may occur with different attitudes or direction of motion.Make the different mode in itself and the library have comparability the standardization of the target exploitation Zelnick square that is partitioned into, and can reduce the workload in structural model storehouse.The Zelnick square just has the character of invariable rotary, but as to make it have translation and yardstick is constant, as long as standardize by regular square.The present invention research be infrared target in remote imaging because operating distance is far away, realizing that identifying purpose process mesoscale do not have significant change, so can in algorithm, omit this step.

5. it is right to determine to be used in the feature extraction layer structural element of the morphology conversion of " hitting miss " according to nearest neighbor method.

Feature extraction layer is that the target after the standardization of Zelnick square is carried out Shape Feature Extraction.Its feature extraction functions utilize morphology " to hit miss " conversion realizes, the right definite method of " hitting miss " mapped structure element that this will relate to a kind of being used for.To target imaging in the empty infrared image, because the phenomenon that target does not have perfect imaging can appear in the influence of factor such as imaging sensor imaging precision and target range be far away.But as long as the target of imaging has due topological structure, promptly inside, target area is smooth, does not exist very for a short time when empty, is to realize Target Recognition by utilizing these incomplete local messages.It is perfect that the present invention utilizes topology information to carry out library, promptly not only deposits complete models in the library, but also deposit the non-complete models that satisfies the topological structure requirement, and arrange according to hierarchical pattern.A kind of integrated pattern in the library derives a group mode class like this.The structural element that utilizes this library can be identified for very easily in " hitting miss " conversion of Shape Feature Extraction is right.

According to the definition in the morphology, for image A, need to determine so a pair of structural element to B=(E, F), a detection image inside, another detection image outside, it is defined as:

A*B＝(AE)∩(A ^CF)

And if only if, and E moves to that certain can insert the inside of A when a bit, and in the time of can inserting A outside when F moves to this, this expression input picture satisfies the shape need of " hitting miss " conversion.Obviously, E should not be connected with F, i.e. E ∩ F=φ.By the previous mode storehouse hierarchy, a certain pattern l in Zelnick square standardization and library is the most approaching when target, then l can be as structural element E, structural element F is then represented by the outline of highest pattern in the l hierarchical pattern class of living in.When practice, after the target exploitation method of principal axis that is partitioned into will being proofreaied and correct usually, carry out " hitting miss " conversion again.

6. normalized result imports the feature extraction layer of invariable rotary neural network, utilizes morphologic " hitting miss " conversion to realize the target shape feature extraction, and will calculate the eigenwert input category network of gained.The process sorter network calculates and finally obtains is the judgement of non-target.

The present invention carried out target to input picture before feature extraction layer cuts apart, and the target that is partitioned into is carried out the standardization of Zelnick square.According to the result after the standard, utilize nearest neighbor method in a different k mode class, to find k corresponding pattern, and can determine that according to this k pattern the structural element in the feature extraction layer is right.Dotted line or dotted line have represented different structural elements to (E in the feature extraction network _i, E _i) (i=1 ... k), all structural elements are to (E _i, F _i) target pattern that is partitioned into is carried out " hitting miss " computing respectively.Like this, for each target to be identified, the Classification and Identification network is the propagated forward network with k input, and the number of different mode class is corresponding in k and the library.

The input of sorter network is to determine according to the transformation results in the feature extraction layer.When object during, get coefficient q=1, otherwise be 0 by " hitting miss " computing.The present invention defines the input of a kind of target signature conspicuousness as sorter network.For recognition system, when object-by shape information to be identified was tending towards complete, at this moment the object conspicuousness that is identified as target also was tending towards 1.Making the area of the most complete pattern in the library is A1, and the target area that is partitioned into is A2, and when q=1, then conspicuousness is defined as: Sa1=A2/A1.When q=0, Sa0=0.Define conspicuousness Sa=q*A2/A1 like this.Network is input as the weighted sum of k conspicuousness value.

The sorter network structure has adopted backpropagation (BP) neural network structure.Network training reaches 1000 circulations or root-mean-square error is lower than at 0.001 o'clock and stops.A large amount of experiments show that network training can reach the root-mean-square error requirement 50 circulations with interior usually, proves that network has fine convergence.

It below is calculated examples at two groups of different training samples.

When A. having than strong correlation at training sample and test samples, invariable rotary morphology neural network recognition effect.So-called sample correlations is meant that by force when target was in the same orientation angle, each was organized and always comprises the sample similar with test samples in the training sample.Select 3 groups of continuous true infrared sequence image samples for use, every group 200 frame amounts to 600 frames.First group, target is marked section transit time by point target to appearance.This stage is mainly detected this paper algorithm and utilizes algorithm of target detection auxiliary positioning target, realizes utilizing side direction part target information (being that target shape information is imperfect) to finish Target Recognition.Second group, the target travel attitude changes, and is used for the recognition performance that detection algorithm changes targeted attitude, and these group data are not used in training network, and all are used for the recognition capability of supervising network.The 3rd group, target is used for detection algorithm and arrives the complete shape information recognition performance of transit time in imperfect shape information along perpendicular to the flight of sensor imaging plane.

Training sample is by first group and the 3rd group extracting 200 frames every frame and form in totally 400 frame data.The node number of network input layer is by the right number decision of structural element.The structural element logarithm is a given constant, determines before network training.Adopt three groups of structural elements to (13 every group) in the experiment, it is right that promptly side direction, forward and crabbing turn to the structural element of transition period of forward flight.Hidden layer is made up of 3 unit.Output layer is made up of 1 unit, and the value of output is between 0 and 1, and having represented the object of identification is the degree of confidence of target or non-target, and degree of confidence is greater than representing to be input as target at 0.5 o'clock.The learning rate of network gets 0.05, and the network training number of times is 1000 times.The all-network weights are got the random number between [0.5,0.5].Fig. 2 is the real infrared image of input, and Fig. 2 (a) belongs to first group of training sample, and Fig. 2 (b), Fig. 2 (c) belong to the 3rd group of training sample, and the former shape is still imperfect, and latter's shape is complete substantially.With the test samples fan-in network, its recognition result sees Table 1 to network through training.

Invariable rotary morphology neural network recognition result during table 1 sample strong correlation

B. have at training sample and test samples more weak when relevant, invariable rotary morphology neural network recognition effect.Two groups of experimental results are when training sample and test samples have strong correlation, and the invariable rotary neural network has good recognition performance.In order to investigate its recognition performance when training sample and test samples correlativity are more weak relatively, this partial design second group of experiment.Last one group of experiment training sample and test samples are to extract every frame from same continuous sequence, and the correlativity between them is very strong.Three groups of samples that correlativity is relatively poor have been adopted in the experiment of this group, and these three groups of sample standard deviations are identical position angles, and promptly target all is the direction flight along the forward imaging plane, but sample is from three mutual discontinuous sequences.200 frame data of training sample, target just in time are in crabbing and turn to forward flight original state.The target flight attitude of test samples one and training sample difference are very big, and stronger noise jamming has appearred in the sample image background of test samples two.Fig. 3 (a), Fig. 3 (b) and Fig. 3 (c) belong to training sample respectively, first group of test samples and second group of test samples.It is right adopt to make that to training sample identification reaches the structural element that satisfies recognition correct rate requirement (90%), and the parameter of sorter network is selected with last one group of experiment identical, and its recognition result sees Table 2.

Invariable rotary morphology neural network recognition result when being correlated with a little less than table 2 sample

By table 1 and table 2 recognition result, no matter the correlativity power of training sample and test samples, invariable rotary morphology neural network recognition effect all can reach very high recognition correct rate, and recognition performance is very stable.

Claims (1)

1, a kind of Infrared Target Recognition Method based on invariable rotary morphology neural network is characterized in that comprising following concrete steps:

1) inner according to the imageable target zone is the smooth empty feature that do not exist, utilize the topology information of pattern to make up library, make and not only deposit complete models in the library, but also deposit the non-complete models that satisfies the topological structure requirement, and, make a kind of integrated pattern in the library derive a group mode class according to the hierarchical pattern layout;

2) according to the imaging characteristics of actual infrared image, and the random character of interframe correlativity, background distributions homogeneity and noise profile during the point target motion, realize that target travel flight path energy accumulation to improve signal noise ratio (snr) of image, realizes the point target location;

3) target position information that provides of combining target detection-phase, realize the local segmentation of target, utilize target in infrared image, to present high luminance area, and surrounding environment belongs to the priori in low-light level district, with target detection point as seed points, set a given threshold value, the connection topological structure constraint that utilizes imageable target to have utilizes the seed growth method, and target is split from background;

4) utilize the rotational invariance of Zelnick square, with the target specificationization that is partitioned into, at first calculate the Zelnick square of a given image, the center of image as true origin, the coordinate Mapping of pixel is in the unit circle scope, and the pixel of those unit circles outside does not participate in calculating, and stores then according to the numerical value of the different orders of Zelnick polynomial computation Zelnick square, and as a n dimensional vector n;

5) the Zelnick square of the various mode class in the library is compared with target Zelnick square, utilize two norm distances of vector, realize the location of target in library in conjunction with nearest neighbor method, and it is right to be identified for the structural element of the morphology conversion of " hitting miss ";

6) normalized result is imported the feature extraction layer of invariable rotary neural network, utilize morphologic " hitting miss " conversion to realize the target shape feature extraction, and will calculate the eigenwert input category network of gained, calculate the judgement that finally obtains " right and wrong " target through sorter network.

Images