CN103295241A - Frequency domain saliency target detection method based on Gabor wavelets - Google Patents

Abstract

The invention relates to a frequency domain saliency target detection method based on Gabor wavelets. The method includes the steps: S1, switching inputted color images into gray images and building a directional feature map; S2, building two color feature maps according to sensitivity of human eyes on different colors of the inputted color images; S3, building a gray feature map for the inputted color images; S4, building a polynomial matrix by the aid of the four feature maps; S5, performing Fourier transformation for the polynomial matrix comprising the four feature maps to obtain a frequency domain polynomial matrix and extracting a magnitude spectrum matrix; S6, performing Gaussian low-pass filter for the magnitude spectrum matrix by a plurality of scales and performing polynomial inverse Fourier transformation for a group of magnitude spectra to obtain a plurality of time domain polynomial matrixes; and S7, dividing the time domain polynomial matrixes into different time domain polynomials according to different scale factors, building a histogram for each time domain polynomial, calculating a one-dimensional entropy function and extracting a time domain saliency map corresponding to the minimum information entropy as a final detection result.

Description

A kind of frequency domain conspicuousness object detection method based on the Gabor small echo

Technical field

The present invention relates to a kind of Gabor of utilization small echo and fourier transform method detects the conspicuousness target, do not utilize the priori of any target image, be a technology of utilizing the psychology of vision model that image is detected, be used for image processing, compression of images, computer vision and target detection track and localization.

Background technology

Visual attention model is a kind ofly to come the model of simulating human visual attention system with computing machine, and extracting human eye institute in piece image can observed attractive focus, for computing machine, is exactly the salient region of this image.Human vision system had both required it to have the ability of handling a large amount of input information, required to have quasi real time reaction capacity again.Psychology of vision studies show that, when the input scene of Analysis of Complex, the human visual system has taked a kind of calculative strategy of serial, namely utilize Selective Attention Mechanism, according to the local feature of image, select the specific region of scene, and by the scanning of eye movement fast, this zone moved on to have high-resolution foveal region of retina, realize the attention to this zone, in order to it is carried out meticulousr observation and analysis.This can regard as the graphical analysis of full visual field and scene understood and finishes by the time-division processing of less partial analysis task.As seen, Selective Attention Mechanism is a human gordian technique selecting particular region of interest from the bulk information of external world's input, with the target detection in the automatic target identification similarity is arranged, therefore, studying the application of human Selective Attention Mechanism in target detection has significance.

At present, the conspicuousness algorithm of target detection mainly is divided into the time domain processing and frequency domain is handled two kinds of thinkings, but two kinds of thinkings all are to come from the computation model that ITTI proposed in 1998, this model is drawn by psychology of vision, and the factor that influences the eye-observation image has been divided into three quantifiable calculated amount---color, gray scale and direction.Along with research deepens continuously, based on time domain conspicuousness object detection method classical conspicuousness method (STB), neural visible sensation method (NVT) etc. are arranged.Consider that the multiple dimensioned process of original calculating is too loaded down with trivial details, the someone proposes to handle conspicuousness model detection algorithm at frequency domain, and spectral residual method (SR), phase spectrum Fourier transform method (PFT) and four-tuple Fourier transform method (QFT) etc. are arranged.

After ITTI was applied to the psychology of vision model target detection of computer vision field in 1998, this method had obtained using widely.Be different from traditional object detection method, its testing process is without any need for priori, that is to say that it does not need earlier Target Photo to be trained modeling, do not need to set up and the initialization sorter yet, it utilizes the information of image self target is understood and to be extracted, make the conspicuousness detection algorithm lack several magnitudes than traditional object detection method computation complexity, also can carry out in real time and be applied in the target following.Though this algorithm can access position and the profile of target, can not identify the attribute of target, the conspicuousness target detection has developed into a new research direction.

Summary of the invention

In order to coordinate computation complexity and conspicuousness accuracy of detection better, the present invention proposes a kind of frequency domain conspicuousness object detection method based on the Gabor small echo.

For realizing such purpose, technical scheme of the present invention comprises following steps:

Step S1: convert input color image to gray level image, utilize the two-dimensional Gabor wavelet filter that gray level image is carried out filtering, set up direction character figure;

Step S2: according to the susceptibility of human eye to the different colours of input color image, set up the color characteristic figure on a basis, the red, green, blue passage is revised, calculate the pixel value of yellow channels then, red, green, blue, the Huang of this moment are four color basis matrixs of color characteristic figure, utilize four color basis matrixs to calculate the poor of the difference of red, green passage and indigo plant, yellow passage, obtain two color characteristic figure and set up two color characteristic figure;

Step S3: utilize input color image to make gray feature figure;

Step S4: utilize direction character figure, two color characteristic figure, gray feature figure to set up polynomial matrix;

Step S5: the polynomial matrix to direction characteristic pattern, two color characteristic figure, gray feature figure compositions carries out Fourier transform, the frequency domain polynomial matrix that obtains; The frequency domain polynomial matrix can be decomposed into amplitude spectrum matrix and phase spectrum matrix, keeps the phase spectrum matrix, extracts the amplitude spectrum matrix;

Step S6: the amplitude spectrum matrix is carried out Gauss's low-pass filtering of a plurality of yardsticks, obtain filtered one group of amplitude spectrum; To one group of amplitude spectrum recycling polynomial expression inverse-Fourier transform, obtain a plurality of time domain polynomial matrix;

Step S7: a plurality of time domain polynomial matrix are divided into different time domain polynomial expressions according to the different scale factor; Each time domain polynomial expression is done histogram; To each histogram calculation one dimension entropy function that obtains; The value of the entropy function minimum of contrast different scale is desired best scale testing result; Extract the remarkable figure of time domain of minimal information entropy correspondence as final detection result.

Beneficial effect of the present invention:

Utilize the Gabor wavelet character to extract and the Fourier transform frequency spectrum analysis method carries out the conspicuousness target detection to image, reach higher detection precision and processing speed faster.

The present invention compares with classic method, and since the vision significance model proposed, it is very fast to have obtained using widely at computer vision field.Because the method that it utilizes imitation human brain visual cortex to handle for target in the visual field finally calculates interested target, thereby allows the same function that can identify interesting target automatically with human eye of computer realization.Technology among the present invention is utilized the Gabor small echo at frequency domain target to be carried out conspicuousness and is handled, and obtains salient region after the inverse transformation, can improve classic method in precision and the arithmetic speed of context of detection.

Utilize yardstick to be σ ∈ { 2 ^-1, 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶One group of gauss low frequency filter the polynomial expression amplitude spectrum of frequency domain is carried out filtering, can be as found from the results, its shirtsleeve operation effect can reach the purpose that conspicuousness detects.Its filtering result is the multiple dimensioned result of conspicuousness target detection.

Because background image often repeats or the consecutive periods appearance, this just makes the amplitude spectrum of its frequency domain behind Fourier transform become impulse function.Utilize the Gabor wavelet character to extract and the Fourier transform frequency spectrum analysis method carries out the conspicuousness target detection to image, reach higher detection precision and processing speed faster.Utilize frequency domain inverse transformation after amplitude spectrum is handled to access target like this and significantly scheme, again in conjunction with multiscale analysis, in eight yardsticks, utilize the two-dimensional entropy function to obtain optimal scale, significantly scheme thereby obtain last target.Conspicuousness object detection method contrast with traditional has fast operation, the characteristics that accuracy is high.

Description of drawings

Fig. 1 is algorithm overall flow figure of the present invention.

The result that Fig. 2 adopts two-dimensional Gabor filter the Digital Image Processing normal pictures to be carried out filtering for the embodiment of the invention.

The result that Fig. 3 adopts two-dimensional Gabor filter that data picture in the present embodiment is carried out filtering for the embodiment of the invention.

The conspicuousness target detection design sketch that Fig. 4 handles natural scene for algorithm of the present invention.

Fig. 5 is the conspicuousness target detection design sketch of algorithm of the present invention to processing such as artificial vehicle instruments.

The conspicuousness target detection design sketch that Fig. 6 handles the people in the natural scene for algorithm of the present invention.

The conspicuousness target detection design sketch that Fig. 7 handles the people in the artificial scene for algorithm of the present invention.

The conspicuousness target detection design sketch that Fig. 8 handles a plurality of natural targets for algorithm of the present invention.

The conspicuousness target detection design sketch that Fig. 9 handles psychology of vision test pattern target for algorithm of the present invention.

Embodiment

Below in conjunction with accompanying drawing embodiments of the invention are elaborated.Present embodiment is being to implement under the prerequisite with the technical solution of the present invention, provided detailed embodiment and concrete operating process, but protection scope of the present invention is not limited to following embodiment.

The realization that present embodiment based target conspicuousness detects, input picture is several natural images and part psychology mode image.

To algorithm overall flow figure, this example provides a kind of computing machine and Gabor small echo and fourier transform method of utilizing that target is carried out the conspicuousness detection, comprises the steps: as shown in Figure 1

Step S1: utilize the Gabor small echo to the extraction of direction characteristic pattern.Because image is a matrix data, what used here is the Gabor wave filter of two dimension, and it can react image orientation information accurately, has the characteristic of simulation eye-observation and extraction target signature.Because natural image is coloured image, we need be translated into gray level image earlier:

$I=\frac{1}{3}(r+g+b)---(1-1)$

I represents the gray level image of input color image in the formula (1-1), and r represents the red pixel passage of input color image, and g represents the green pixel passage of input color image, and b represents the blue pixel passage of input color image.

The two-dimensional Gabor filter size of using among the present invention is 11 * 11 pixel square, and it is described below:

$G(x,y)=\exp (-\frac{(X^2+{\mathrm{\γ}}^2+Y^2)}{{2\mathrm{\σ}}^2})\cos \left(\frac{2\mathrm{\π}}{\mathrm{\λ}}X\right)---(1-2)$

X＝xcosθ-ysinθ (1-3)

Y＝xsinθ+ycosθ (1-4)

G represents the Gabor matrix of two dimension in the formula (1-2), the row-coordinate of x represent pixel, and the row coordinate of y represent pixel, and x, y ∈ 5 ..., 5}.X and Y are the coordinate transforms after the adding angle parameter theta, are provided by formula (1-3) and formula (1-4) respectively.Also have three constants in addition, scale factor γ value in the present invention is 0.3, and effective width σ value in the present invention is 4.5, and wavelength X value in the present invention is 5.6.

Utilizing formula (1-2) that gray level image is carried out filtering handles, described two-dimensional Gabor filter size is 11 * 11 pixel square, choose the angle θ of 4 different directions={ 0 ° in the present invention, 45 °, 90 °, 135 ° } carry out the filtering of four angles, get local maximum in conjunction with the filtering result of four angles then and obtain direction character figure, four angles are carried out after the filtering and got the image that maximum filtering result obtains.Adopt two-dimensional Gabor filter that the Digital Image Processing normal pictures is carried out the result of filtering and the result that Fig. 3 adopts two-dimensional Gabor filter that data picture in the present embodiment is carried out filtering as Fig. 2, we carry out after the filtering to 0 ° of digital picture, 45 °, 90 °, 135 ° four angle digital pictures and data picture and get the image that maximum filtering result obtains, and the method for solving of maximum filtering matrix O is: the pixel of four filtered direction matrix correspondence positions is got maximal value and obtained.

Step S2: according to the susceptibility of human eye to the different colours of input color image, set up the color characteristic figure on a basis, the red, green, blue passage is revised, calculate the pixel value of yellow channels then, red, green, blue, the Huang of this moment are four color basis matrixs of color characteristic figure, utilize four color basis matrixs to calculate the poor of the difference of red, green passage and indigo plant, yellow passage, obtain two color characteristic figure and set up two color characteristic figure; The specific descriptions of the foundation of color characteristic figure are as follows:

$R=r-\frac{(g+b)}{2}---(2-1)$

$G=g-\frac{(r+b)}{2}---(2-2)$

$B=b-\frac{(r+g)}{2}---(2-3)$

$Y=\frac{(r+g)}{2}-\frac{|r-g|}{2}-b---(2-4)$

Formula (2-1) is in the formula (2-4), and r, g, b are identical with meaning in the formula (1-1), and R, G, B, Y then are four color basis matrixs of color characteristic figure, represent red, green, blue, Huang respectively.Therefore we can access two color characteristic figure, by:

RG＝R-G (2-5)

BY＝B-Y (2-6)

Formula (2-5) and (2-6) in R, G, B, Y obtained to (2-4) by formula (2-1), two matrixes of RG and BY are exactly two color characteristic figure among the present invention.

Step S3: set up gray feature figure.Because main detection mode is to carry out at frequency domain, the gray feature figure I that the gray feature figure among the present invention directly utilizes formula (1-1) to obtain.

Step S4: utilize direction character figure, two color characteristic figure, gray feature figure to set up polynomial matrix; According to the synthetic polynomial matrix of characteristic pattern.Because each feature differs for the contribution of the expression of conspicuousness target, so can not simple addition handle, therefore set up polynomial matrix Q for four characteristic patterns that extract above and be expressed as follows:

$Q=\mathrm{aRG}+\mathrm{bBY}\stackrel{\→}{x}+\mathrm{cI}\stackrel{\→}{y}+\mathrm{dO}\stackrel{\→}{z}---(4-1)$

It is formula (2-5) and result of calculation (2-6) that formula (4-1) middle RG, BY represent two color characteristic figure, and I represents gray feature figure, and O represents direction character figure.Wherein a, b, c, d represent polynomial constant coefficient, and we get a=b=c=d=0.25 in the present invention. Then being polynomial base vector, also is direction vector.

Step S5: the polynomial matrix to direction characteristic pattern, two color characteristic figure, gray feature figure compositions carries out Fourier transform, the frequency domain polynomial matrix that obtains; The frequency domain polynomial matrix can be decomposed into amplitude spectrum matrix and phase spectrum matrix, keeps the phase spectrum matrix, extracts the amplitude spectrum matrix; Polynomial expression Fourier transform: because the background object that frequency domain repeats for processing has good classification character, so the polynomial matrix that we form characteristic pattern carries out Fourier transform, extract amplitude spectrum then the conspicuousness target is extracted.Polynomial expression Fourier transform f (n m) is expressed as follows:

f(n，m)＝a+bi+cj+dk (5-1)

Suppose our polynomial expression suc as formula shown in (5-1), n, m represent discrete row-coordinate and row coordinate respectively, and i, j, k are representing polynomial base vector.Therefore, this polynomial expression Fourier transform is expressed as:

$F_H[u,v]=\frac{1}{\sqrt{\mathrm{MN}}}\underset{m=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}{e}^{-\mathrm{\μ}2\mathrm{\π}(\left(\frac{\mathrm{mv}}{M}\right)+\left(\frac{\mathrm{nu}}{N}\right))}f(n,m)---(5-2)$

U, v represent the two-dimensional coordinate of frequency domain, F in the formula (5-2) _HRepresent the frequency domain polynomial matrix behind the Fourier transform, (n m) is obtained by formula (5-1) f, represents the time domain polynomial matrix, and M, N represent length and the width of matrix respectively, and μ represents imaginary part unit, i.e. μ ²=-1.

Step S6: the amplitude spectrum matrix is carried out Gauss's low-pass filtering of a plurality of yardsticks, obtain filtered one group of amplitude spectrum; To one group of amplitude spectrum recycling polynomial expression inverse-Fourier transform, obtain a plurality of time domain polynomial matrix; Amplitude spectrum filtering.Because target context generally has the characteristic that the cycle repeats, sky for example, meadow, road etc.The shock response of the corresponding frequency domain of unlimited periodic signal in the time domain, and time domain limit cycle signal, corresponding the pulse signal of frequency domain.Therefore show as the form of pulse signal in the frequency domain amplitude spectrum of target context behind the polynomial expression Fourier transform, utilize simple low pass filter just can finish the inhibition of pulse signals like this, thereby finish the eliminating to target context, finally finish the purpose to the conspicuousness target detection.

Here provide the gauss low frequency filter form of using among the present invention:

$H(u,v)=e^{-\frac{D^2(u,v)}{{2\mathrm{\σ}}^2}}---(6-1)$

H is electric-wave filter matrix in the formula (6-1), and the D representative is used Euclidean distance apart from the distance of Fourier transform initial point among the present invention.σ represents the degree of expansion of Gaussian curve.In order to consider the yardstick unchangeability, the degree of expansion σ of Gaussian curve has got 8 different values and has carried out the filtering of different scale.We get σ ∈ { 2 among the present invention ^-1, 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶, utilize yardstick to be σ ∈ { 2 ^-1, 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴⁵, 2 ⁶One group of gauss low frequency filter the polynomial expression amplitude spectrum of frequency domain is carried out filtering, the filtering result is the multiple dimensioned result of conspicuousness target detection.

We make, and polynomial matrix Q is Q through the frequency domain polynomial matrix that obtains after the polynomial expression Fourier transform _H, its amplitude spectrum is expressed as so:

A＝|Q _H| (6-2)

A represents the amplitude spectrum of frequency domain polynomial matrix in the formula (6-2).Amplitude is carried out the gaussian filtering of 8 yardsticks:

A _H＝A×H (6-3)

A in the formula (6-3) _HRepresent filtered one group of amplitude spectrum.Recycling polynomial expression inverse-Fourier transform obtains the time domain polynomial matrix, and inverse transformation is expressed as follows:

$f(n,m)=\frac{1}{\sqrt{\mathrm{MN}}}\underset{v=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{u=0}{\overset{N-1}{\mathrm{\Σ}}}{e}^{\mathrm{\μ}2\mathrm{\π}(\left(\frac{\mathrm{mv}}{M}\right)+\left(\frac{\mathrm{nu}}{N}\right))}{F}_H[u,v]---(6-4)$

The implication of each parameter in the formula (6-4) is identical with formula (5-2).

Step S7: a plurality of time domain polynomial matrix are divided into different time domain polynomial expressions according to the different scale factor; Each polynomial expression is done histogram; To each histogram calculation one dimension entropy function that obtains; The value of the entropy function minimum of contrast different scale is our desired best scale testing result; Extract the remarkable figure of time domain of minimal information entropy correspondence as final detection result.

Multiscale analysis.Owing to need select only yardstick as last remarkable figure, we think that best testing result should have abundanter visual information, therefore entropy function is expanded to two dimension, utilize the two-dimensional entropy function to come multiple dimensioned remarkable figure is selected, choose our last remarkable figure of conduct of entropy function value minimum.

Calculate the two-dimensional entropy function, earlier the remarkable figure of a plurality of yardsticks that obtains is done histogram, obtain the probability of each pixel by histogram, by histogrammic conversion, the Pixel Information that is equivalent to the image of two dimension has been become one dimension is handled.

Fig. 4 has showed the result that the conspicuousness of algorithm process picture of the present invention detects to Fig. 8, wherein comprises the scenery of natural scene according to the classification of type of target, the people of natural scene, the vehicle of natural scene, the people of artificial scene etc.Wherein the number classification according to target comprises that single goal, binocular are marked with and multiple goal.More show the conspicuousness detection figure of algorithm of the present invention when handling psychology of vision test picture among Fig. 9, can see the notice concentrated area that still meets very much the eye-observation picture.

The present invention obtains the four direction characteristic pattern after with filtering, asks local maximum to obtain direction character figure then, and it is representing the contour feature of image.Color combining characteristic pattern and gray feature figure carry out polynomial Fourier transform then, take out amplitude spectrum, and it is used low-pass Gaussian filter filtering, finish the purpose that conspicuousness detects.Because background image often repeats or the consecutive periods appearance, this just makes the amplitude spectrum of its frequency domain behind Fourier transform become impulse function.Utilize frequency domain inverse transformation after amplitude spectrum is handled to access target like this and significantly scheme, again in conjunction with multiscale analysis, in eight yardsticks, utilize the two-dimensional entropy function to obtain optimal scale, significantly scheme thereby obtain last target.Conspicuousness object detection method contrast with traditional has fast operation, the characteristics that accuracy is high.

The above; only be the embodiment among the present invention, but protection scope of the present invention is not limited thereto, anyly is familiar with the people of this technology in the disclosed technical scope of the present invention; conversion or the replacement expected can be understood, all of the present invention comprising within the scope should be encompassed in.

Claims (4)

1. frequency domain conspicuousness object detection method based on the Gabor small echo is characterized in that comprising that concrete steps are as follows:

Step S1: convert input color image to gray level image, utilize the two-dimensional Gabor wavelet filter that gray level image is carried out filtering, set up direction character figure;

Step S2: according to the susceptibility of human eye to the different colours of input color image, set up the color characteristic figure on a basis, the red, green, blue passage is revised, calculate the pixel value of yellow channels then, red, green, blue, the Huang of this moment are four color basis matrixs of color characteristic figure, utilize four color basis matrixs to calculate the poor of the difference of red, green passage and indigo plant, yellow passage, set up two color characteristic figure;

Step S3: input color image is made gray feature figure;

Step S4: utilize direction character figure, two color characteristic figure, gray feature figure to set up polynomial matrix;

Step S5: the polynomial matrix to direction characteristic pattern, two color characteristic figure, gray feature figure compositions carries out Fourier transform, the frequency domain polynomial matrix that obtains; The frequency domain polynomial matrix can be decomposed into amplitude spectrum matrix and phase spectrum matrix, keeps the phase spectrum matrix, extracts the amplitude spectrum matrix;

Step S6: the amplitude spectrum matrix is carried out Gauss's low-pass filtering of a plurality of yardsticks, obtain filtered one group of amplitude spectrum; To one group of amplitude spectrum recycling polynomial expression inverse-Fourier transform, obtain a plurality of time domain polynomial matrix;

Step S7: a plurality of time domain polynomial matrix are divided into different time domain polynomial expressions according to the different scale factor; Each time domain polynomial expression is done histogram; To each histogram calculation one dimension entropy function that obtains; The value of the entropy function minimum of contrast different scale is desired best scale testing result; Extract the remarkable figure of time domain of minimal information entropy correspondence as final detection result.

2. the frequency domain conspicuousness object detection method based on the Gabor small echo as claimed in claim 1, it is characterized in that: described two-dimensional Gabor filter size is 11 * 11 pixel square, get angle θ={ 0 °, 45 °, 90 °, 135 ° } carry out the filtering of four angles, get local maximum in conjunction with the filtering result of four angles then and obtain direction character figure, four angles are carried out after the filtering and got the image that maximum filtering result obtains.

3. the frequency domain conspicuousness object detection method based on the Gabor small echo as claimed in claim 1, it is characterized in that: described polynomial matrix Q is expressed as follows:

$Q=\mathrm{aRG}+\mathrm{bBY}\stackrel{\→}{x}+\mathrm{cI}\stackrel{\→}{y}+\mathrm{dO}\stackrel{\→}{z}$

RG, BY represent two color characteristic figure in the formula, and I represents gray feature figure, and O represents direction character figure; Wherein a, b, c, d represent polynomial constant coefficient,

Then being polynomial base vector, also is direction vector.

4. the frequency domain conspicuousness object detection method based on the Gabor small echo as claimed in claim 1 is characterized in that: utilize yardstick to be σ ∈ { 2 ^-1, 2 ⁰, 2 ¹, 2 ², 2 ³, 2 ⁴, 2 ⁵, 2 ⁶One group of gauss low frequency filter the polynomial expression amplitude spectrum of frequency domain is carried out filtering, the filtering result is the multiple dimensioned result of conspicuousness target detection.

Images