CN106548457A - A kind of method for detecting infrared puniness target using multi-direction first-order partial derivative - Google Patents

Abstract

The present invention relates to a kind of method for detecting infrared puniness target using multi-direction first-order partial derivative, it has three big steps：First, the binary cubic function of a small range gradation of image being constructed using Cubic facet model, the first-order partial derivative of all directions, and the convolution mask of design coefficient being solved in central area, each operation coefficient can be obtained in image directly from after convolution；Then, the first-order partial derivative characteristic according to Small object in all directions, strengthens convolution mask by the design of inner product maximization principle, in all directions target is strengthened；Finally, multiplication fusion is carried out to the result in individual all directions, suppresses background while target is further enhanced as much as possible, obtain end product.The Dim targets detection of infrared image is the composite can be widely applied to, with wide market prospect and using value.

Description

A kind of method for detecting infrared puniness target using multi-direction first-order partial derivative

(1) technical field

The present invention relates to a kind of method for detecting infrared puniness target using multi-direction first-order partial derivative, belongs to digital picture Process field, relates generally to Cubic facet model and target detection technique.There is wide answering in all kinds of application systems based on image Use prospect.

(2) background technology

The detection of Weak target plays the effect of key, sky in infraed early warning system due to image-forming range farther out The targets such as the steamer under aircraft, Sea background under background would generally become a less object, find mesh in advance exactly Mark can take appropriate measures in advance and carry out and prevent or dispose.But make an uproar as IR Scene has substantial amounts of sensor Sound and heterogeneity interference, along with various mixed and disorderly strong edges etc. under detection scene itself, therefore can cause false alarm rate It is too high, how to suppress false-alarm, always one important problem while effective detection target.Based on maximum-medium filtering With the method for maximum-mean filter (referring to document：Di Shipandemeng etc., the maximum-intermediate value and maximum for small target deteection- Mean filter, SPIE's light is studied science, engineering and instrument international symposium collection of thesis, and 1999:74-83. (Deshpande S D,Meng H E,Venkateswarlu R,et al.Max-mean and max-median filters for detection of small targets[C]//SPIE's International Symposium on Optical Science,Engineering,and Instrumentation.International Society for Optics and Photonics,1999:74-83.)) it is a kind of classical small target detecting method, by selecting some under current pixel neighborhood The intermediate value or average of specific orientation positions filters Small object substituting the pixel of current operation with this, but for white Gaussian Noise is more sensitive, easily causes false-alarm, and the filtering method with it as representative there is also identical problem.Some are based on form Classical small target detecting method is (referring to document：Bai Xiangzhi etc., new top cap are converted and its in small IR targets detection application In analysis and research, pattern recognition, 2010：43(6)：2145-2156.(Bai X,Zhou F.Analysis of new top- hat transformation and the application for infrared dim small target detection[J].Pattern Recognition,2010,43(6):2145-2156.)) strengthen target using top cap conversion Suppress background, morphology operations are simple and quick, but when complex scene is processed, easily disturbed so as to false-alarm by strong edge etc. It is too high, while the size of morphological operator is also fixed mostly, it is impossible to while adaptively being adjusted according to scene.In recent years, some Some effects are achieved (referring to document based on the detection method of small target of rarefaction representation：High Chen Qiang etc., in single image The infrared block image model of detection Small object, IEEE's image procossing transactions, 2013,22 (12): 4996-5009.(Gao C,Meng D,Yang Y,et al.Infrared patch-image model for small target detection in a single image[J].IEEE Transactions on Image Processing, 2013,22(12):4996-5009.)), but due to the uncertainty of Small object distribution so that when design object function Need to consider various situations, cause amount of calculation excessively complicated, while model lacks explanatory, effect is also limited.Small and weak mesh , due to the salience of its intensity profile, some are using the method for this diversity (referring to document for mark：Deng He etc., is weighed based on local The infrared small target detection method of weight difference measurement, IEEE's geography and remote sensing transactions, 2016,54 (7):4204-4214.(Deng H,Sun X,Liu M,et al.Small Infrared Target Detection Based on Weighted Local Difference Measure[J].IEEE Transactions on Geoscience and Remote Sensing,2016,54(7):4204-4214.)) by designing a kind of tolerance so that the tolerance of target area can be with Background area is different from significantly, so as to realize detection, but as IR Scene complexity is various, some simple measure of criterions are difficult To distinguish target and background in detail, therefore effect is limited.In order to strengthen the robustness of algorithm, some algorithms make use of multi-direction (referring to document：Qi Shengxiang etc., a kind of infrared small target detection side based on robust direction significance under environment complicated and changeable Method, IEEE's geography and remote sensing bulletin, 2013,10 (3):495-499.(Qi S,Ma J,Tao C, et al.A robust directional saliency-based method for infrared small-target detection under various complex backgrounds[J].IEEE Geoscience and Remote Sensing Letters,2013,10(3):495-499.)) and multi-scale information is (referring to document：Poplar length ability etc., for infrared The multiple dimensioned Cubic facet model of small target deteection, Infrared Physics ＆ Technology, 2014,67:202-209.(Yang C,Ma J,Zhang M,et al.Multiscale facet model for infrared small target detection[J] .Infrared Physics&Technology,2014,67:202-209.)) and achieve certain effect, it is in general red Small object quantity of information in outer scene is few, lack texture and shared pixel is few, therefore directly can using its indexing distributed intelligence Certain limitation can be had, it is considered to how to carry out strengthening under other angles remaining a problem for being worth exploring.

The Dim targets detection of infrared image is more difficult, and main thought is strengthened by entering line translation to original image, So as to realize the detection of Small object, while in order to be applied to actual scene, also having to the complexity and run time of algorithm Higher requirement.In order to realize quickly and efficiently detecting Weak target, by the first derivative information of a diffusion model, this Invention proposes a kind of method for detecting infrared puniness target using multi-direction first-order partial derivative.

(3) content of the invention

1st, purpose：Dim targets detection be infrared early warning with guidance system in important step, but existing detection method Target can not effectively be detected.Traditional all kinds of algorithms, may be due to complexity while preferable verification and measurement ratio is pursued Edge or noise jamming under environment causes false-alarm too high.

In order to solve the above problems and make up the deficiency of traditional method, the invention provides a kind of inclined using multi-direction single order The method for detecting infrared puniness target of derivative, it analyzes diffusion model single order in all directions from Cubic facet model Derivative Characteristics, and strengthened in different directions according to the isotropism of target area respectively, in order to increase as efficiently as possible Strong target, devises enhancing convolution mask according to inner product maximization principle, finally obtains the result multiplication fusion in all directions Last result figure.Most of process can be obtained by the mask convolution for having designed, and improved accuracy of detection, reduced empty There is obvious improvement in alert rate and run time.

2nd, technical scheme：In order to realize this purpose, technical scheme is as follows, first, using Cubic facet model structure The binary cubic function of a small range gradation of image is made, and the first-order partial derivative of all directions, and design is solved in central area The convolution mask of coefficient, each operation coefficient can be obtained in image directly from after convolution；Then, according to Small object in each side First-order partial derivative characteristic upwards, strengthens convolution mask by the design of inner product maximization principle, in all directions target is entered Row strengthens；Finally, multiplication fusion is carried out to the result in individual all directions, is suppressed while target is further enhanced as much as possible Background, obtains end product.

The present invention relates to a kind of method for detecting infrared puniness target using multi-direction first-order partial derivative, the method is specifically walked It is rapid as follows：Step one：The binary cubic function of a small range gradation of image is constructed using Cubic facet model, is solved in central area each The first-order partial derivative in direction, and the convolution mask of required coefficient is devised, the figure that each operation coefficient can directly from after convolution Obtain as in.

Polynomial equation of the Cubic facet model by least square fitting a small range, discrete gray value is transformed to continuously Functional value, with this accurate derivative value for solving each point, while with certain noise immunity.

Define two set R={-2-1 012 } and C={-2-1 012 }, three times in the range of set R × C Toroidal function can be made up of following Discrete Orthogonal base： Using S as the symmetric neighborhood in the range of R × C, Ir, c are current point gray value, then cubic surface function can To be expressed as：

Wherein, by least square fitting, each coefficient k_iCan be expressed as follows, g_iFor i-th polynomial orthogonal base：

Therefore, each coefficient is considered as mask convolution operation and obtains, and Current central point (0, it is level 0), vertical and each First derivative on the α of direction is：

For k₂,k₃,k₇,k₈,k₉And k₁₀If their corresponding convolution masks are w₂,w₃,w₇,w₈,w₉With w₁₀, by the above Formula can be calculated, and each convolution mask design is as follows：

w₃=w₂ ^T

w₁₀=w₇ ^T

w₈=w₉ ^T

Step 2：First-order partial derivative characteristic according to Small object in all directions, is increased by the design of inner product maximization principle Strong convolution mask, strengthens to target in all directions：

We establish actual point diffusion model (see Fig. 2), and the first-order partial derivative that four direction is carried out to which is solved.Figure 3 list level, the derivative distribution of vertical and two clinodiagonals and its two dimension view, it can be seen that obey point diffusion profile Small object in all directions have extremely similar rule, therefore consider strengthened in all directions respectively.

After original image is solved first derivative in all directions, 4 sides are separately designed in order to effectively strengthen Small object Wave filter upwards.As the general occupied area of target is less, the effective volume of actual distribution is restricted, therefore considers 7 Design in the range of × 7 strengthens convolution mask.

The process of mask convolution can be regarded as the inner product of two high dimension vectors, just possess intensity profile due to there was only target Isotropism, only need to strengthen target area in all directions as far as possible, you can while suppressing background.By the target of all directions Regard a dimensional vector V as in region_t, convolution strengthen template regard a dimensional vector V as_c, by the unit vector on its directionWith Vectorial mould k is multiplied, and the amount for needing design isBy angle formulae：

Obtain：

K=1 is made, works as V_tWhen fixed, in order that target inner product maximum is equivalent to the maximum for seeking cos θ, the i.e. size of inner product The cosine value of angle is proportional to, so convolution mask should will be strengthenedIt is designed as similar to all directions first derivative of target Shape, k in practice can be chosen according to practical situation.By taking horizontal direction as an example, process is chosen as shown in figure 4, remaining direction In the same manner.Step 3：Multiplication fusion is carried out to the result in individual all directions, is suppressed while target is further enhanced as much as possible Background, obtains end product.

The enhancing figure in all directions, last result figure f have been obtained in being located at the first two steps_resultIt is represented by：

f_result=f₀·f₄₅·f₉₀·f₁₃₅

3rd, advantage and effect：A kind of method for detecting infrared puniness target using multi-direction first-order partial derivative of the present invention, from Cubic facet model sets out, and analyzes diffusion model first derivative characteristic in all directions, and according to each to same of target area Property is strengthened respectively in different directions, in order to strengthen target as efficiently as possible, is devised according to inner product maximization principle Strengthen convolution mask, finally the result in all directions is multiplied to merge obtains last result figure.Most of process can be by The mask convolution for having designed is obtained, and is significantly changed improving accuracy of detection, reducing having in false alarm rate and run time It is kind, all kinds of application systems based on image are can be widely applied to, with wide market prospect and using value.

(4) illustrate

Fig. 1 is a kind of theory diagram of the method for detecting infrared puniness target using multi-direction first-order partial derivative of the present invention.

Fig. 2 is the point diffusion model of present invention analysis.

Fig. 3 is the four direction first derivative of midpoint diffusion model of the present invention.

Fig. 4 is the design process that the present invention strengthens convolution mask in all directions.

Fig. 5 is testing result of the present invention in actual scene, and wherein Fig. 5 a- Fig. 5 i are original images, and Small object is by white Color square frame labelling, Fig. 5 a '-Fig. 5 i ' are last testing results.

Fig. 6 a～Fig. 6 i are the ROC curve comparative result of the detection method with other several methods of the present invention.

(5) specific embodiment

Embodiments of the present invention are made further below in conjunction with accompanying drawing by technical scheme for a better understanding of the present invention Description.

A kind of method for detecting infrared puniness target using multi-direction first-order partial derivative of the present invention, theory diagram such as Fig. 1 institutes Show, specific implementation step is as follows：

Step one：The binary cubic function of a small range gradation of image is constructed using Cubic facet model, is solved in central area The first-order partial derivative of all directions, and the convolution mask of required coefficient is devised, each operation coefficient can directly from after convolution Obtain in image.

Polynomial equation of the Cubic facet model by least square fitting a small range, discrete gray value is transformed to continuously Functional value, with this accurate derivative value for solving each point, while with certain noise immunity.

Define two set R={-2-1 012 } and C={-2-1 012 }, three times in the range of set R × C Toroidal function can be made up of following Discrete Orthogonal base： Using S as the symmetric neighborhood in the range of R × C, Ir, c are current point gray value, then cubic surface function can To be expressed as：

Wherein, by least square fitting, each coefficient k_iCan be expressed as follows, g_iFor i-th polynomial orthogonal base：

Therefore, each coefficient is considered as mask convolution operation and obtains, and Current central point (0, it is level 0), vertical and each First derivative on the α of direction is：

For k₂,k₃,k₇,k₈,k₉And k₁₀If their corresponding convolution masks are w₂,w₃,w₇,w₈,w₉With w₁₀, by the above Formula can be calculated, and each convolution mask design is as follows：

w₃=w₂ ^T

w₁₀=w₇ ^T

w₈=w₉ ^T

Step 2：First-order partial derivative characteristic according to Small object in all directions, is increased by the design of inner product maximization principle Strong convolution mask, strengthens to target in all directions.

The present invention establishes actual point diffusion model (see Fig. 2), and the first-order partial derivative that four direction is carried out to which is solved. Fig. 3 lists level, the derivative distribution of vertical and two clinodiagonals and its two dimension view, it can be seen that obey point diffusion minute The Small object of cloth has extremely similar rule in all directions, therefore considers to be strengthened in all directions respectively.

After original image is solved first derivative in all directions, 4 sides are separately designed in order to effectively strengthen Small object Wave filter upwards.As the general occupied area of target is less, the effective volume of actual distribution is restricted, therefore considers 7 Design in the range of × 7 strengthens convolution mask.

The process of mask convolution can be regarded as the inner product of two high dimension vectors, just possess intensity profile due to there was only target Isotropism, only need to strengthen target area in all directions as far as possible, you can while suppressing background.By the target of all directions Regard a dimensional vector V as in region_t, convolution strengthen template regard a dimensional vector V as_c, by the unit vector on its directionWith Vectorial mould k is multiplied, and the amount for needing design isBy angle formulae：

Obtain：

K=1 is made, works as V_tWhen fixed, in order that target inner product maximum is equivalent to the maximum for seeking cos θ, the i.e. size of inner product The cosine value of angle is proportional to, so convolution mask should will be strengthenedIt is designed as similar to all directions first derivative of target Shape, k in practice can be chosen according to practical situation.By taking horizontal direction as an example, process is chosen as shown in figure 4, remaining direction In the same manner.Step 3：Multiplication fusion is carried out to the result in individual all directions, is suppressed while target is further enhanced as much as possible Background, obtains end product.

The enhancing figure in all directions, last result figure f have been obtained in being located at the first two steps_resultIt is represented by：

f_result=f₀·f₄₅·f₉₀·f₁₃₅

The actual point diffusion model that Fig. 2 is set up for easy analysis by the present invention.Fig. 3 is the model four sides First derivative distribution upwards, it can be seen that its distribution in all directions has closely similar rule.Fig. 4 is by vector Product maximization principle, devises the enhancing convolution mask in all directions on the premise of known target vector.Fig. 5 is the present invention Application in actual IR Scene, in the original infrared image of Fig. 5 a- Fig. 5 i, the position of Small object is marked with white edge, figure 5a '-Fig. 5 i ' are corresponding testing result.Fig. 6 a～Fig. 6 i are detection method and other several detection methods in the present invention ROC curve comparison diagram, in the multiple image that Fig. 5 is listed, the ROC curve this method under each scene achieves best Effect.

Image for experiment comes from different IR Scenes, and Small object great majority therein are very dim, in noise Increase the difficulty of detection under environment again, but experimental result not only restrained effectively background enhanced target, realize fast Speed is detected exactly, and also has obvious advantage in the contrast with additive method, and this absolutely proves that the present invention's is effective Property, and the detecting system of all kinds of infrared small objects is can be widely applied to, with wide market prospect and using value.

Claims (1)

1. a kind of method for detecting infrared puniness target using multi-direction first-order partial derivative, it is characterised in that：The method is specifically walked It is rapid as follows：

Step one：Two set R={-2-1 012 } and C={-2-1 012 } are defined, from Cubic facet model, in collection Close the cubic surface function in the range of R × C to be made up of following Discrete Orthogonal base： Using S as the symmetric neighborhood in the range of R × C, then toroidal function is expressed as follows：

$\begin{array}{c}f(r,c)=k_1+k_{2}r+k_{3}c+k_4(r^2-2)+k_{5}rc+k_6(c^2-2)+k_7(r^3-\frac{17}{5}r)+k_8(r^2-2)c\\ +k_{9}r(c^2-2)+k_{10}\left(c^3-\frac{17}{5}c\right)\end{array}$

The level of Current central point (0,0), the first derivative vertically and in all directions α are：

$\frac{\∂f}{\∂r}=k_2-\frac{17}{5}{k}_7-2k_9$

$\frac{\∂f}{\∂c}=k_3-\frac{17}{5}{k}_{10}-2k_8$

$f_{\mathrm{\α}}^{\′}=\frac{\∂f}{\∂r}sin\mathrm{\α}+\frac{\∂f}{\∂c}cos\mathrm{\α}$

For k₂,k₃,k₇,k₈,k₉And k₁₀, their corresponding convolution masks are designed for w by method of least square₂,w₃,w₇,w₈,w₉ With w₁₀, each multinomial coefficient directly can be obtained by convolution operation；

Step 2：First-order partial derivative characteristic according to Small object in all directions, strengthens volume by the design of inner product maximization principle Product module plate, strengthens to target in all directions：

After original image is solved first derivative in all directions, the wave filter on 4 directions is separately designed；Due to target one As occupied area it is less, the effective volume of actual distribution is restricted, therefore considers that design strengthens convolution mould in the range of 7 × 7 Plate；The process of convolution can be regarded as the inner product of two high dimension vectors, by the target area of all directions regard as it is one-dimensional arrange to Amount V_t, convolution strengthen template regard a dimensional vector V as_c, by the unit vector on its directionIt is multiplied with vectorial mould k, needs to set The amount of meter isBy angle formulae：

$cos\mathrm{\&theta;}=\frac{<V_t,V_c>}{\left|V_t\right|\left|V_c\right|}$

Obtain：

$<V_t,k\stackrel{\&RightArrow;}{e}>=\left|V_t\right|kcos\mathrm{\&theta;}$

K=1 is made, works as V_tWhen fixed, in order that target inner product is maximum, it is equivalent to seek the maximum of cos θ, so convolution should will be strengthened TemplateThe shape similar to all directions first derivative of target is designed as, k in practice can be chosen according to practical situation；Step Rapid three：Multiplication fusion is carried out to the result in individual all directions, is suppressed background while target is further enhanced as much as possible, is obtained To end product：

The enhancing figure in all directions, last result figure f have been obtained in being located at the first two steps_resultIt is represented by：

f_result=f₀·f₄₅·f₉₀·f₁₃₅。