CN103345730B - Based on the Infrared Image Processing Method of lateral inhibition network - Google Patents

Abstract

The present invention relates to a kind of Infrared Image Processing Method based on lateral inhibition network, the method specifically comprises the following steps S1: obtain the local entropy in each neighborhood of pixel points of image, obtains the local entropy matrix of image; S2: according to entire image local entropy matrix, obtain matrix maximal value, then structure tolerance operator; S3: solve rejection coefficient matrix; S4: process infrared image according to step S2 and S3 acquired results, obtains lateral inhibition network and exports.The method is according to image saltation zone (comprising edge and details) and smooth area (comprising background and the noise) different qualities in spatial domain, utilize Local Entropy of Image to construct one to measure operator, lateral inhibition network is made to have limb recognition ability, noise immunity is strong, and in conjunction with filtering mechanism, while enhancing image, also there is inhibiting effect to noise.

Description

Based on the Infrared Image Processing Method of lateral inhibition network

Technical field

The present invention relates to infrared image enhancement technical field, be specifically related to a kind of Infrared Image Processing Method based on lateral inhibition network.

Background technology

Vision lateral inhibition is that Hartline and colleague thereof are doing king crab compound eye to find in electro physiology experimentation the earliest.Years of researches result shows, human visual system also exists lateral inhibition function, and this comes from the inhibiting effect between adjacent Visual Neuron, and at the pretreatment stage of visual signal, lateral inhibition effect is considered to play vital effect.

Infrared image is the infrared emanation energy imaging by detector receiving target, therefore compared with visible images, infrared image has that contrast is low, edge fog and the feature such as signal to noise ratio (S/N ratio) is low, and lateral inhibition network is actually the Hi-pass filter in a spatial frequency, one of topmost function is exactly outstanding frame, enhanced contrast, therefore can process infrared image with lateral inhibition, image high-frequency information is enhanced, and low-frequency information is inhibited simultaneously, thus improves contrast.But because infrared image inevitably introduces noise, noise and edge details belong to high frequency, lateral inhibition network is while enhancing edge details, the noise of image is made to have also been obtained amplification, thus have impact on visual effect, also be unfavorable for the subsequent treatment of image, therefore how utilize lateral inhibition network restraint speckle while enhancing image to be the emphasis studied in the last few years.

The innovatory algorithm of current Contralateral suppression network mainly can be divided into two classes: a class is improving one's methods based on transform domain, the most frequently used transform domain is the method based on frequency domain, specifically in frequency domain, distinguish image high-frequency information and low-frequency information by Fourier transform exactly, then process respectively, the image that spatial domain is enhanced is changed in last contravariant, and these class methods have good treatment effect, but calculated amount is larger, can not meet real time handling requirement, therefore practicality is not high.Another kind of is improving one's methods based on spatial domain, concrete is exactly the feature in spatial domain according to picture noise and marginal information, process in conjunction with filtering mechanism, these class methods have the advantages such as algorithm is simple, processing speed is fast, real-time, are current topmost methods.

Xie Xiaofang, the people such as Mao Xiaobo are in " the lateral inhibition network algorithm for image enhancement research of improvement " literary composition, analyze the difference of image local neighborhood gray-scale value and noise information, propose the method that a kind of auto adapted filtering mechanism combines with lateral inhibition network, the output of self-adaptative adjustment wave filter can be carried out according to the variance of the average of the local neighborhood of image and variance and noise, although this method inhibits noise while strengthening image to a certain extent, but need to know noise variance information in advance, and noise variance information is unknown often in actual conditions, therefore there is larger limitation.

Known by analyzing above, the image edge details after existing lateral inhibition network process is enhanced, and contrast is improved, but noise information have also been obtained amplification.The innovatory algorithm calculated amount of some lateral inhibition networks is large, is unfavorable for real-time process, and limitation is comparatively large, is difficult to be applied in a practical situation.

Summary of the invention

In view of this, the object of this invention is to provide a kind of limb recognition ability strong, noise immunity is strong, while enhancing image, also has the inhibiting Infrared Image Processing Method based on lateral inhibition network to noise,

The object of the invention is by such technical scheme realize,

Based on an Infrared Image Processing Method for lateral inhibition network, comprise the following steps:

S1: obtain the local entropy in each neighborhood of pixel points of image, obtains the local entropy matrix of image, obtains entire image local entropy matrix further;

S2: according to entire image local entropy matrix, obtain matrix maximal value, then structure tolerance operator;

S3: generate lateral inhibition matrix of coefficients;

S4: infrared image is processed according to step S2 and S3 acquired results.

Further, described step S1 specifically comprises following sub-step:

S11: the neighborhood getting N × N size in image space centered by pixel (m, n), counts the number of times that in neighborhood, each gray level occurs, obtains the probability P that each gray level occurs _j; in formula: P _jfor the probability that gray level j occurs in pixel (m, n) neighborhood, m _jfor having the total pixel number of gray level j in neighborhood;

S12: obtain local entropy H in this neighborhood of pixel points: in formula, L represents that maximum gray scale is other;

S13: obtain the local entropy in each neighborhood of pixel points of image, obtain Local Entropy of Image matrix H.

Further, the value of the other L of described maximum gray scale is 256.

Further, described step S2 specifically comprises following sub-step:

S21: the maximal value H obtaining Local Entropy of Image matrix H _max:

H _max＝max(H)

S22: obtain the gray average that pixel (m, n) is neighborhood with radius l' size

$\stackrel{\‾}{F(m,n)}=\frac{l^{\′}}{{(2\×l^{\′}+1)}^2}\underset{k1=-l^{\′}}{\overset{l^{\′}}{\mathrm{\Σ}}}\underset{k2=-l^{\′}}{\overset{l^{\′}}{\mathrm{\Σ}}}F(k1,k2),$

In formula: k1, k2 represent each pixel coordinate in pixel (m, n) radius l ' neighborhood;

S23: solve tolerance operator E, its computing formula is:

$E=\frac{H(m,n)}{H_{\max }}(F(m,n)-\stackrel{\‾}{F(m,n)})$

In formula: H (m, n) represents the local entropy of pixel (m, n), F (m, n) represents the gray-scale value of input image pixels point (m, n), represent pixel (m, n) neighboring mean value.

Further, in described step S3, select bimodal Gauss model to solve lateral inhibition matrix of coefficients, then lateral inhibition coefficient is expressed as: $k_{\mathrm{mn},\mathrm{pq}}=\frac{1}{\mathrm{\β}}[\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_1}\exp (-\frac{d_{\mathrm{mn},\mathrm{pq}}^2}{{2\mathrm{\σ}}_1^2})-\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_2}\exp (-\frac{d_{\mathrm{mn},\mathrm{pq}}^2}{{2\mathrm{\σ}}_2^2})],$ β, σ in formula ₁, σ ₂be all constant, represent central pixel point (m, n) and the Euclidean distance of pixel (p, q), select suitable neighborhood, obtain lateral inhibition matrix of coefficients.

Further, in described step S4, select the output of acyclic network calculation side Suppression network, the output of lateral inhibition network is expressed as: $G(m,n)=\stackrel{\‾}{F(m,n)}-\underset{p=-l}{\overset{l}{\mathrm{\Σ}}}\underset{q=-l}{\overset{l}{\mathrm{\Σ}}}{k}_{\mathrm{mn},\mathrm{pq}}F(m+p,n+q)+\mathrm{\λE},$ Wherein k _{mn, pq}for lateral inhibition coefficient, λ is one and regulates constant, and l is lateral inhibition radius.

Further, described λ span is [0,1].

Owing to have employed technique scheme, the present invention has following advantage:

The present invention proposes the Infrared Image Processing Method based on lateral inhibition network, with gray average in pixel (m, n) neighborhood in lateral inhibition network input replace F (m, n), make lateral inhibition network have certain filter action.Tolerance operator E is added in lateral inhibition network, network area can be made to separate edge, evenly point and noise spot, different inhibition strengths is taked to these points, and high brightness point relative brightness on edge can be made larger, the relative brightness of low-light level point is less, such marginal point contrast is more obvious, and also inhibit noise to a certain extent, that filtering mechanism can also be avoided to bring is fuzzy simultaneously.Also noise spot may be there is in edge in addition, therefore while strengthening edge, also enhance the noise on edge, but human eye will be far smaller than smooth area for noise-sensitive degree on edge, from this point the present invention structure tolerance operator meet human-eye visual characteristic.

Other advantages of the present invention, target and feature will be set forth to a certain extent in the following description, and to a certain extent, based on will be apparent to those skilled in the art to investigating hereafter, or can be instructed from the practice of the present invention.Target of the present invention and other advantages can be realized by instructions below and obtain.

Accompanying drawing explanation

In order to make the object, technical solutions and advantages of the present invention clearly, below in conjunction with accompanying drawing, the present invention is described in further detail, wherein:

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is original image;

Fig. 3 is the local entropy image corresponding with Fig. 2;

Fig. 4 is the local entropy value that in Fig. 2, black line is capable;

Fig. 5 is that lateral inhibition network strengthens image;

Fig. 6 is that the lateral inhibition network improved strengthens image.

Embodiment

Below with reference to accompanying drawing, the preferred embodiments of the present invention are described in detail; Should be appreciated that preferred embodiment only in order to the present invention is described, instead of in order to limit the scope of the invention.

As shown in Figure 1, a kind of Infrared Image Processing Method based on lateral inhibition network, comprises the following steps: the local entropy of image solves; Structure tolerance operator; Solve lateral inhibition matrix of coefficients; Lateral inhibition network is used for image procossing.

Entropy is mainly used in the uncertainty representing things state, image entropy reflect image containing quantity of information number, be exactly specifically the space characteristics representing gradation of image distribution.In edge details district, neighborhood territory pixel gray level is more, then local entropy is comparatively large, and illustrate containing quantity of information many, background area neighborhood gray level is less, and local entropy is less, few containing quantity of information.If get the neighborhood of N × N size at the aerial image vegetarian refreshments (m, n) of image, so tried to achieve entropy is exactly the local entropy of pixel (m, n).

S1: obtain the local entropy in each neighborhood of pixel points of image, obtains the local entropy matrix of image, obtains entire image local entropy matrix further;

S11: the neighborhood getting N × N size in image space centered by pixel (m, n), counts the number of times that in neighborhood, each gray level occurs, obtains the probability P that each gray level occurs _j; in formula: P _jfor the probability that gray level j occurs in pixel (m, n) neighborhood, m _jfor having the total pixel number of gray level j in neighborhood;

S12: obtain local entropy H in this neighborhood of pixel points:

$H=-\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}{P}_j\log P_j$

In formula, L represents that maximum gray scale is other; The value of general L is 256;

S13: travel through this image, obtains the local entropy in each neighborhood of pixel points of image, obtains Local Entropy of Image matrix H.

Local entropy is only relevant with image local grey scale change degree, and with local gray-value change size have nothing to do, and infrared image edge gray-value variation is often very little, less at some weak edges, the advantage of its uniqueness is therefore distributed with local entropy reflection infrared image local gray level.

S2: according to entire image local entropy matrix, obtain matrix maximal value, then structure tolerance operator;

S21: the maximal value H obtaining Local Entropy of Image matrix H _max:

H _max＝max(H)

S22: obtain the gray average that pixel (m, n) is neighborhood with radius l' size

$\stackrel{\‾}{F(m,n)}=\frac{l^{\′}}{{(2\×l^{\′}+1)}^2}\underset{k1=-l^{\′}}{\overset{l^{\′}}{\mathrm{\Σ}}}\underset{k2=-l^{\′}}{\overset{l^{\′}}{\mathrm{\Σ}}}F(k1,k2),$

S23: solve tolerance operator E, its computing formula is:

$E=\frac{H(m,n)}{H_{\max }}(F(m,n)-\stackrel{\‾}{F(m,n)})$

In formula: H (m, n) represents the local entropy of pixel (m, n), F (m, n) represents the gray-scale value of input image pixels point (m, n), represent pixel (m, n) neighboring mean value.

For marginal point, H (m, n) is very large, absolute value also comparatively large, therefore in equation, E absolute value is very large.It is less for smooth area noise spot H (m, n), comparatively large, therefore E is larger.For the even point of smooth area, H (m, n) is less, less, therefore E is minimum.If point (m, n) is high brightness point, then e be larger on the occasion of; If point (m, n) is low-light level point, then e is larger negative value.So tolerance operator has just distinguished edge, noise and background.

S3: generate lateral inhibition matrix of coefficients;

Experiment shows, the inhibiting effect of the suffered receptor unit around of arbitrary receptor unit reduces along with the increase of the spacing of receptor unit, and therefore lateral inhibition coefficient can regard the function of distance d as.Current lateral inhibition Modulus Model mainly contains: hyperbolic model, Gauss model, bimodal Gauss model, exponential model.These models all meet the trend of monotone decreasing in designation area, are specifically inversely proportional to regard to the distance reflected between inhibiting intensity and receptor unit in lateral inhibition, and namely distance is larger, and inhibition strength is less.

Research shows, have the receptor of a segment distance to be also eager to excel than the most contiguous ommatidium effect from tested receptor, therefore bimodal Gauss model is closest to vision actual conditions, then lateral inhibition coefficient is expressed as:

$k_{\mathrm{mn},\mathrm{pq}}=\frac{1}{\mathrm{\β}}[\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_1}\exp (-\frac{d_{\mathrm{mn},\mathrm{pq}}^2}{{2\mathrm{\σ}}_1^2})-\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_2}\exp (-\frac{d_{\mathrm{mn},\mathrm{pq}}^2}{{2\mathrm{\σ}}_2^2})]$

β, σ in formula ₁, σ ₂be all constant, represent central pixel point (m, n) and the Euclidean distance of pixel (p, q), select suitable neighborhood, obtain lateral inhibition matrix of coefficients.

S4: infrared image is processed according to step S2 and S3 acquired results.

Lateral inhibition mechanism be due to visual receptor unit between interaction produce, it is a kind of mode of vision system process information, on the one hand space contrast information is enhanced, on the other hand the space information of duplicating is inhibited, thus enhance edge details, what fuzzy image was become is clear.Lateral inhibition network model has many, wherein be Hartline-Ratliff model the most widely, be divided into circulation side Suppression network and acyclic lateral inhibition network, wherein circulation side Suppression network calculated amount is very large, and be difficult to the real-time process accomplishing image, therefore we select acyclic network, and add tolerance operator in a network, then by the rejection coefficient matrix of trying to achieve and each pixel convolution of image, carry out infrared image processing, the output of the lateral inhibition network finally obtained can be expressed as:

$G(m,n)=\stackrel{\‾}{F(m,n)}-\underset{p=-l}{\overset{l}{\mathrm{\Σ}}}\underset{q=-l}{\overset{l}{\mathrm{\Σ}}}{k}_{\mathrm{mn},\mathrm{pq}}F(m+p,n+q)+\mathrm{\λE}$

Wherein k _{mn, pq}for lateral inhibition coefficient, λ is one and regulates constant, and general λ span is [0,1], which increase the dirigibility of lateral inhibition network process image, l is lateral inhibition radius, in order to make treatment effect better, average radius l ' can get different values from lateral inhibition radius l.

Fig. 2 represents original image, and Fig. 3 is the local entropy image (mark that to be people be adds of the black line in figure is capable) of the former figure obtained.As seen from the figure, the local entropy of image border point is greater than background area, and Fig. 4 gives in Fig. 3 and draws the capable local entropy value of black line, can find out, although the noise of background area causes impact to local entropy, but still be less than marginarium.

By the image that Fig. 5 is the enhancing of former lateral inhibition network, while can finding out that image is enhanced, noise have also been obtained amplification, and visual effect is bad.Fig. 6 is the image that the present invention strengthens, and while image enhaucament, noise have also been obtained suppression, better than former lateral inhibition network treatment effect.

The present invention analyzes the deficiency of existing lateral inhibition network, proposes a kind of lateral inhibition network and improves one's methods, with gray average in pixel (m, n) neighborhood in lateral inhibition network input replace F (m, n), make lateral inhibition network have certain filter action.Tolerance operator E is added in lateral inhibition network, network area can be made to separate edge, evenly point and noise spot, different inhibition strengths is taked to these points, and high brightness point relative brightness on edge can be made larger, the relative brightness of low-light level point is less, such marginal point contrast is more obvious, and also inhibit noise to a certain extent, that filtering mechanism can also be avoided to bring is fuzzy simultaneously.The present invention is that a kind of lateral inhibition network is improved one's methods, and has theoretical and practical value preferably to infrared image enhancement technology.

The foregoing is only the preferred embodiments of the present invention, be not limited to the present invention, obviously, those skilled in the art can carry out various change and modification to the present invention and not depart from the spirit and scope of the present invention.Like this, if these amendments of the present invention and modification belong within the scope of the claims in the present invention and equivalent technologies thereof, then the present invention is also intended to comprise these change and modification.

Claims (6)

1. based on an Infrared Image Processing Method for lateral inhibition network, it is characterized in that: comprise the following steps:

S1: obtain the local entropy in each neighborhood of pixel points of image, obtains the local entropy matrix of image, obtains entire image local entropy matrix further;

S2: according to entire image local entropy matrix, obtain matrix maximal value, then structure tolerance operator;

S3: generate lateral inhibition matrix of coefficients;

S4: infrared image is processed according to step S2 and S3 acquired results;

Described step S2 specifically comprises following sub-step:

S21: the maximal value H obtaining Local Entropy of Image matrix H _max:

H _max＝max(H)

S22: obtain the gray average that pixel (m, n) is neighborhood with radius l' size

$\stackrel{\‾}{F(m,n)}=\frac{l^{\′}}{{(2\×l^{\′}+1)}^2}\underset{k1=-l^{\′}}{\overset{l}{\Σ}}\underset{k2=-l^{\′}}{\overset{l}{\Σ}}F(k1,k2),$

In formula: k1, k2 represent each pixel coordinate in pixel (m, n) radius l ' neighborhood;

S23: solve tolerance operator E, its computing formula is:

$E=\frac{H(m,n)}{H_{max}}(F(m,n)-\stackrel{\‾}{F(m,n)})$

In formula: H (m, n) represent pixel ( _m,n) local entropy, F (m, n) represents the gray-scale value of input image pixels point (m, n), represent pixel (m, n) neighboring mean value.

2. a kind of Infrared Image Processing Method based on lateral inhibition network according to claim 1, is characterized in that: described step S1 specifically comprises following sub-step:

S11: the neighborhood getting N × N size in image space centered by pixel (m, n), counts the number of times that in neighborhood, each gray level occurs, obtains the probability P that each gray level occurs _j;

In formula: P _jfor the probability that gray level j occurs in pixel (m, n) neighborhood, m _jfor having the total pixel number of gray level j in neighborhood;

S12: obtain local entropy H in this neighborhood of pixel points: in formula, L represents that maximum gray scale is other;

S13: obtain the local entropy in each neighborhood of pixel points of image, obtain Local Entropy of Image matrix H.

3. a kind of Infrared Image Processing Method based on lateral inhibition network according to claim 2, is characterized in that: the value of the other L of described maximum gray scale is 256.

4. a kind of Infrared Image Processing Method based on lateral inhibition network according to claim 1, is characterized in that: in described step S3, and select bimodal Gauss model to solve lateral inhibition matrix of coefficients, then lateral inhibition coefficient is expressed as:

$k_{mn,pq}=\frac{1}{\mathrm{\β}}\[\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_1}\exp (-\frac{d_{mn,pq}^2}{2{\mathrm{\σ}}_1^2})-\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_2}\exp (-\frac{d_{mn,pq}^2}{2{\mathrm{\σ}}_2^2})\],$

β, σ in formula ₁, σ ₂be all constant, represent central pixel point (m, n) and the Euclidean distance of pixel (p, q), select suitable neighborhood, obtain lateral inhibition matrix of coefficients.

5. a kind of Infrared Image Processing Method based on lateral inhibition network according to claim 4, is characterized in that: in described step S4, and select the output of acyclic network calculation side Suppression network, then the output of lateral inhibition network is expressed as: $G(m,n)=\stackrel{\‾}{F(m,n)}-\underset{p=-l}{\overset{l}{\Σ}}\underset{q=-l}{\overset{l}{\Σ}}{k}_{mn,pq}F(m+p,n+q)+\mathrm{\λ}E,$ Wherein k _{mn, pq}for lateral inhibition coefficient, λ is one and regulates constant, and l is lateral inhibition radius.

6. a kind of Infrared Image Processing Method based on lateral inhibition network according to claim 5, is characterized in that: described λ span is [0,1].