CN108154147A - The region of interest area detecting method of view-based access control model attention model - Google Patents
The region of interest area detecting method of view-based access control model attention model Download PDFInfo
- Publication number
- CN108154147A CN108154147A CN201810034712.3A CN201810034712A CN108154147A CN 108154147 A CN108154147 A CN 108154147A CN 201810034712 A CN201810034712 A CN 201810034712A CN 108154147 A CN108154147 A CN 108154147A
- Authority
- CN
- China
- Prior art keywords
- image
- scale
- notable
- gabor
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
Abstract
The present invention relates to a kind of region of interest area detecting methods of view-based access control model attention model, are related to technical field of image processing.The present invention is improved Itti vision modes, add the visual perception process of vision noticing mechanism simulation people, increase and play target movable information so that testing result more meets the physiological characteristic of people, is significantly improved to the detection result of Missile Body mark class Small object under complex background.
Description
Technical field
The present invention relates to technical field of image processing, and in particular to a kind of area-of-interest inspection of view-based access control model attention model
Survey method.
Background technology
People are often put to interested object more note that this is because vision attention is selective, and claim
For visual selective or vision significance.When in face of complex scene, vision system can be primarily focused in scene rapidly
Certain marking areas.
Region of interest (Regions Of Interest, ROI) is people's observation and generates interest, concern when understanding image
Or the region paid attention to, i.e., it most can be intriguing in image, it can most show the region of picture material.In the application of its attention mechanism
The key concept of proposition.Region of interest may be considered in image the most notable pixel set of (saliency), i.e., significant point or
The set of point of interest.How region of interest is automatically extracted from piece image, be exactly region of interest detection technique.
Region of interest detect, i.e., salient region detection is using computer technology simulate human visual system, using regarding
Feel attention model, extract some key messages of image as significant point, the appropriate area using centered on significant point is emerging as sense
Interesting area.Area-of-interest selects not relying on the unique characteristics of scene areas, but relies on its generation compared with peripheral region
Relative characteristic, i.e. vision significance.Conspicuousness is strong, is chosen as area-of-interest.How the region of interest inspection of image a kind of is designed
Survey method so that the physiological characteristic that testing result more meets people becomes the technical issues of urgently to be resolved hurrily.
Invention content
(1) technical problems to be solved
The technical problem to be solved by the present invention is to：How the region of interest detection method of image a kind of is designed so that detection
As a result more meet the physiological characteristic of people.
(2) technical solution
In order to solve the above technical problem, the present invention provides a kind of area-of-interest detections of view-based access control model attention model
Method includes the following steps：
Step 1: establish multi-scale image structure
One width two dimensional image, scale space under different scale represent can by input picture I (x, y) and Gaussian kernel G (x,
Y, σ) convolution obtains：
L (x, y, σ)=G (x, y, σ) * I (x, y) (2)
In formula (2), the location of pixels of (x, y) representative image, σ is known as the scale space factor, and value is smaller, characterizes the image quilt
Smooth is fewer, and corresponding scale is also just smaller, and large scale corresponds to the general picture feature of image, and small scale corresponds to the thin of image
Feature is saved, L represents the scale space of image；
Step 2: low-level visual features are extracted
1st, brightness extracts
If color video frequency image, r, g, b represent red, green and blue in image respectively, then brightness calculation is public
Formula is：
I=(r+g+b)/3 (3)
If gray level image, then gray feature directly chooses the gray value of each pixel；2nd, color feature extracted
It is respectively red, green, blue, yellow 4 Color Channels to define R, G, B, Y, then：
R=r- (g+b)/2
G=g- (r+b)/2 (4)
B=b- (r+g)/2
Y=(r+g)/2- (r-g)/2-b
Then RG and BY channels are：RG=R-GB, Y=B- (R+G)/2；
3rd, Directional feature extraction
Direction character is extracted with Gabor filter：
One-dimensional Gabor functions, i.e. 1D-Gabor functions：
In formula, standard deviations of the σ for Gaussian function, w0For the spatial frequency of complex plane wave, x0Center for 1D-Gabor functions
Point coordinates, the formula of the odd, even component of 1D-Gabor functions are respectively：
According to 3 σ principles, l=6 σ/w are taken0；
Two-dimensional Gabor function, i.e. 2D-Gabor functions are one under the Gauss envelopes multiple change sine waves along x-axis, two
Tieing up Gabor function expressions is：
The real and imaginary parts of two-dimensional Gabor function are respectively：
θ is directioin parameter, and σ is bigger, and energy more disperses, smaller more to concentrate, and has high frequency σh, low frequency σlTwo centre frequencies；
The selection of deflection θ：θ=0,30,60,90,120,150, filtering window 32*32 share 24 Gabor filters
Wave device；
Input picture I (x, y) and the convolution of Gabor wavelet kernel function are：R (x, y)=∫ ∫ I (ε, η) g (x- ε, y- η) d ε d
η (8)
Gabor wavelet transformation after the result is that plural number, take plural number mould | | r (x, y) | | mean value and variance as small echo change
The result changed；
θ=0 is selected, the Gabor filter output of 30,60,90,120,150 six directions passes through as direction character
(7) formula obtains the Gaobr wave filters of six direction, and then each tomographic image of gaussian pyramid structure is filtered with these wave filters
Wave obtains the Feature Mapping figure on corresponding six direction；
4th, Motion feature extraction
Motion feature includes movement velocity and two category feature of the direction of motion；
1) motion vector extracts：The motion vector of target is extracted using background subtraction, by calculating each pixel in image
The relevance in time domain or spatial domain between point using the background of former frame as reference, calculates the background of present frame, then by present frame
Image and its background subtracting, obtain difference image, if t frames image is I (x, y, t), corresponding background image is B (x, y, t),
Then difference image is：
2) Motion feature extraction：According to the difference image, obtain the movement of each pixel in adjacent two field pictures away from
From DijThe move distance of each pixel on 0 °, 45 °, 90 ° and 135 ° four direction is projected, then obtains each pixel four by (x, y)
Move distance in a direction of motion, and then the movement velocity for obtaining 0 °, 45 °, 90 ° and 135 ° four direction is followed successively by：
D0°(i, j)=pixel is in the component in x directions
D90°(i, j)=pixel is in the component in y directions
Step 3: characteristic pattern is calculated with notable figure
The center of visual attention model-periphery operation is to carry out interlayer phase reducing to central stratum and peripheral tier, that is, is calculated
The difference of the corresponding large and small scale of different resolution, the mathematic(al) representation of characteristic pattern E are：
E (c, s)=| E (c) Θ E (s) | (11)
Wherein, Θ represents that the interpolation of two different scale images is subtracted each other, i.e., first to peripheral tier into row interpolation, pixel number is increased
The difference operation of central stratum and peripheral tier is carried out after to corresponding central stratum pixel number, c represents center scale, behalf neighboring area
Scale, c ∈ { 2,3,4 }, i.e., 2,3,4 layers are central stratum, s=c+ δ, scale difference δ ∈ { 3,4 }, i.e., each central stratum corresponding week
Boundary layer adds 3 or 4 for the number of plies of central stratum；
The characteristic pattern of gray scale, color, direction and motion feature is acquired according to formula (11)：
Gray feature figure：
I (c, s)=| I (c) Θ I (s) | (12)
Color characteristic figure：RG (c, s)=| (R (c)-G (c)) Θ (G (s)-R (s)) |
BY (c, s)=| (B (c)-Y (c)) Θ (Y (s)-B (s)) | (13)
Direction character figure：
O (c, s, θ)=| O (c, θ) Θ O (s, θ) | (14)
Motion feature figure：Ds(c, s)=| Ds(c)ΘDs(s)|
D0°(c, s)=| D0°(c)ΘD0°(s)|
D45°(c, s)=| D45°(c)ΘD45°(s)| (15)
D90°(c, s)=| D90°(c)ΘD90°(s)|
D135°(c, s)=| D135°(c)ΘD135°(s)|
All kinds of characteristic patterns are merged using default normalization operator N (), obtain the notable figure of each category feature.
Gray scale notable figure：
Color notable figure：
Direction notable figure：
Movement velocity notable figure：
Direction of motion notable figure：
⊕ is default blending algorithm, and characteristic remarkable picture further normalizes, and addition obtains comprehensive notable figure：
After obtaining the synthesis notable figure of image, each position competes with one another in comprehensive notable figure, and the position of triumph, which becomes, to be paid attention to
Focus forms area-of-interest.
Preferably, it in step 3, will successively select in the victor is a king in comprehensive notable figure input Itti vision modes network
The region of conspicuousness maximum in notable figure is selected as area-of-interest.
Preferably, it is further included after step 3 and focus-of-attention is realized by the inhibition of return mechanism in Itti vision modes
The step of transfer.
Preferably, w0=1.
(3) advantageous effect
The present invention is improved Itti vision modes, adds the visual perception process of vision noticing mechanism simulation people,
Increase and play target movable information so that testing result more meets the physiological characteristic of people, to the small mesh of Missile Body mark class under complex background
Target detection result is significantly improved.
Description of the drawings
Fig. 1 is the complicated ground image detection result figure based on Itti visual attention models；Wherein (a) is artwork, and (b) is
The synthesis notable figure of visual attention model, (c) are vision attention testing results；
Fig. 2 is the improvement visual attention model schematic diagram that the present invention uses；
Fig. 3 is the multi-scale image structure diagram that the present invention uses；
Fig. 4 is the one-dimensional Gabor functional digraphs of the present invention；
Fig. 5 is the two-dimensional Gabor functional digraph of the present invention；
Fig. 6 is that the size of the present invention is 4 × 6 Gabor filter group.
Specific embodiment
To make the purpose of the present invention, content and advantage clearer, with reference to the accompanying drawings and examples, to the present invention's
Specific embodiment is described in further detail.
In numerous visual attention models, Itti proposes that visual attention model is of greatest concern.It belongs to bottom-up
Attention mechanism carries out Multiresolution Decomposition after being filtered to linearity, extract brightness, color and the direction character of image as regarding
Feel primary features, image office is calculated using center-periphery (Center-Surround, C-S) calculus of differences and non-linear normalizing
The characteristic attribute (brightness, color and direction) of portion's range obtains the characteristic pattern under different scale, each characteristic pattern is carried out multiple dimensioned
The notable figure of each feature dimensions is obtained after fusion and normalization.The mode of nonuniform sampling is then based on, is merged using multi-characteristic
The notable figure that strategy ties up these different characteristics merges, and forms final synthesis notable figure.It is finally that comprehensive notable figure is defeated
Enter in " the victor is a king " (winner-take-all WTA) network, select successively the region of conspicuousness maximum in notable figure as
Area-of-interest.Attention mobility is carried out using the neural network and " inhibition of return " mechanism of " the victor is a king ".The model can utilize
Human eye is to the attention mechanisms of the features such as color of image, brightness and direction, the adaptively area-of-interest in detection image, in mesh
Have great advantage in mark detection.
A more complicated ground scene image is selected, as shown in Fig. 1 (a).When human eye watches the image attentively, can most draw
Play human eye attention is a building building of image, it belongs to man-made target, is typical ground assault target.Due to ground
Scape is complicated, is difficult accurately to detect the building with conventional dividing method.Utilize the region of interest based on Itti visual attention models
Domain detection algorithm can then detect building position well, and (b) is comprehensive notable figure in Fig. 1, and (c) is testing result, wherein face
Color is redder to represent that attention value is higher.
According to the imaging characteristic of tank target, in high s/n ratio, tank target has brightness, color and texture spy
Sign, can directly be detected it using Itti visual attention models.
According to Missile Body target imaging characteristic, bullet mark has certain brightness, but color and textural characteristics unobvious, directly
It connects and utilizes Itti visual attention models, detection result is undesirable.Below as the modeling approach of Itti vision modes, to model into
Row improves, and increases Missile Body target kinetic characteristic, the improvement visual attention model for being suitble to the detection of Missile Body mark is established, such as Fig. 2 institutes
Show.Specifically, detection method includes the following steps for the area-of-interest of the view-based access control model attention model of the present invention：
Step 1: establish multi-scale image structure
The sampling density and vision addressability of human visual system is as the distance to retinal centre increases and reduces, i.e.,
The process of human eye understanding things is ten split-phase of process of multi-resolution decomposition process and vision system understanding things from thick to thin
Seemingly, thus can analog vision attention model this nonuniform sampling mechanism well.Nonuniform sampling is earliest by Kronauer＆
Zeevi (1985) is proposed, is then frequently used (Rybak 1998, Itti2005) in the model of simulation human vision.Its
In most it is representative be Itti＆Koch (2001) propose gaussian pyramid model, as shown in Figure 3.
Image Multiscale structure describes original image using a series of word image of different resolutions, wherein subgraph at different levels
(or coefficient) represents the input picture that resolution ratio reduces step by step in succession.The formation of multi-resolution framework is to use to carry out certainly image
Decomposition of the bottom to top, each tomographic image are that its previous tomographic image is formed by certain template convolution.Multi-resolution framework can
To efficiently perform many basic image operations, one group of low pass or band logical image can be generated.It is more by the interconnection of grade and grade
Resolution structural provides contacting between Local treatment and Global treatment.
Koendetink proves that Gaussian convolution core is the unique translation core for realizing change of scale, and Gaussian kernel is unique line
Property core.
Two-dimensional Gaussian function is defined as follows：
σ represents the standard deviation of Gauss normal distribution, and x, y are independent variable.
One width two dimensional image, the scale space expression under different scale can be by input picture I (x, y) and Gauss nuclear convolution
It obtains：
L (x, y, σ)=G (x, y, σ) * I (x, y)
(2) in formula (2), the location of pixels of (x, y) representative image, σ is known as the scale space factor, and the smaller then characterization of value should
Image is smoothed fewer, and corresponding scale is also just smaller.Large scale corresponds to the general picture feature of image, and small scale corresponds to figure
The minutia of picture.L represents the scale space of image.
Step 2: low-level visual features are extracted
Human visual system is classified and is identified by capturing clarification of objective.Therefore, the spy of target is extracted
Sign, is the basic premise for establishing visual attention model.Clarification of objective generally comprises color, brightness, direction, size, shape, fortune
Dynamic, solid degree, curvature, topological attribute, closure etc., wherein vision is most sensitive to motion feature.Therefore the present invention is in Itti
On the basis of model, for Missile Body mark vision-based detection, motion feature is introduced, and special with color characteristic, brightness, direction
Sign is together as low-level visual features.
1st, brightness extracts
If color video frequency image, r, g, b represent red, green and blue in image respectively.Then general brightness
Calculation formula is：
I=(r+g+b)/3 (3)
If gray level image, then gray feature directly chooses the gray value of each pixel.2nd, color feature extracted
It is respectively red, green, blue, yellow 4 Color Channels to define R, G, B, Y, then：
R=r- (g+b)/2
G=g- (r+b)/2 (4)
B=b- (r+g)/2
Y=(r+g)/2- (r-g)/2-b
Then RG and BY channels are：RG=R-GB, Y=B- (R+G)/2.
3rd, Directional feature extraction
Direction character is extracted with Gabor filter.One key property of Gabor transformation is that Gabor becomes
The regularity of distribution that coefficient discloses the local frequencies of a signal or piece image is changed, rather than Fourier transformation coefficients only
Reflect the information of global frequencies.Gabor transformation is highly useful in peacekeeping 2D signal processing, and advantage is multiple
Research direction is proved to, such as speech recognition, the detection of signal, compression of images, texture analysis, the segmentation of image and identification side
Face.Another interesting characteristics of Gabor transformation are that Gabor basic functions have the property similar to human vision primitive, especially
It is secondly this function waveform of Wiki, similar to the impression surface wave shape of mammalian visual systems, this is just mathematically dynamic for research
The spatial character in object visual experience face provides a kind of effective mode.Pass through spectrum analysis, it is known that Gabor functions are on frequency domain
It is a bandpass filter, by adjusting the direction of Gabor functions, target can be extracted different towards in upper assigned frequency band
Information, and other unconcerned information are all filtered out, therefore, it is good property detector.
It is a kind of integrated approach based on gray scale and feature based that Gabor wavelet feature, which describes method,.It says it is based on spy
Sign, be because the Gabor wavelet coefficient of extraction different directions and scale is as feature in being distributed from gradation of image, with one group
Feature and its location expression target signature are obtained by optimization, do not need to priori.Say it is based on gray scale be because
For using Gabor wavelet extract feature can complete reconstruction image, other than the mean value of image, do not lose position, shape
Information.
One-dimensional Gabor functions, i.e. 1D-Gabor functions：
In formula, standard deviations of the σ for Gaussian function, w0For the spatial frequency of complex plane wave, x0Center for 1D-Gabor functions
Point coordinates takes w0The formula of the odd, even component of=1,1D-Gabor function is respectively：
According to 3 σ principles, l=6 σ/w are taken0。
Fig. 4 is the figure of the one-dimensional odd, even component of Gabor functions.
Two-dimensional Gabor function, i.e. 2D-Gabor functions are one under the Gauss envelopes multiple change sine waves along x-axis, such as
Shown in Fig. 5.Two-dimensional Gabor function expression is：
The real and imaginary parts of two-dimensional Gabor function are respectively：
There are three the parameters for determining Gabor：f,θ,σ.F determines scale, and θ is directioin parameter, and σ is bigger, and energy more disperses, more
It is small more to concentrate, usually there is high frequency σh, low frequency σlTwo centre frequencies.
The selection of deflection θ：The efficient transformation section that can learn θ is [0,2 π], since the Fourier transformation of image exists
There is symmetry, so the value range of θ is [0, π] in frequency domain.
It is assumed that f=4, θ=0,30,60,90,120,150, filtering window 32*32, share 24 Gabor filters,
Fig. 6 gives the spatial domain expression figure of this group of wave filter.In figure, line direction indicates 6 directions of wave filter from 0 to 150
Variation, column direction illustrate the variation of 4 frequencies of wave filter from high to low.
Input picture I (x, y) and the convolution of Gabor wavelet kernel function are：R (x, y)=∫ ∫ I (ε, η) g (x- ε, y- η) d ε d
η (10)
Gabor wavelet transformation after the result is that plural number, take plural number mould | | r (x, y) | | mean value and variance as small echo change
The result changed.
As can be seen that with the continuous raising of filter frequencies, the feature for the image that wave filter can characterize significantly increases,
I.e. its response in Gabor domains also constantly becomes larger.And as can see from Figure 6 image it is filtered after its direction spy
Sign is significantly characterized.Therefore, either from the angle of frequency still from the angle in direction, by Gabor filter
After group filtering, apparent response of the feature in Gabor wavelet domain is all higher in image, and unconspicuous feature is filtered
Its response is then correspondingly very low after wave.Simultaneously as the advantageous characteristic of Gabor filter in itself, such as scale invariability, rotation
Invariance etc. but also target is passing through certain translation, rotates, after dimensional variation, response of the corresponding points in Gabor domains
Value equally has constant characteristic, consequently facilitating the target with fixed characteristic point is described using these features.
The present invention selects θ=0, and the Gabor filter output of 30,60,90,120,150 six directions is used as direction character,
The Gaobr wave filters of six direction can be obtained by (7) formula, then with these wave filters to each of gaussian pyramid structure
Tomographic image filters, you can obtains the Feature Mapping figure on corresponding six direction.
4th, Motion feature extraction
Motion feature mainly includes movement velocity and two category feature of the direction of motion.
1) motion vector extracts.The motion vector of target is extracted using background subtraction.By calculating each pixel in image
The relevance in time domain or spatial domain between point using the background of former frame as reference, calculates the background of present frame.Then by present frame
Image and its background subtracting, obtain difference image.If t frames image is I (x, y, t), corresponding background image is B (x, y, t),
Then difference image is：
2) Motion feature extraction.According to the difference image, the fortune of each pixel in adjacent two field pictures can be obtained
Dynamic distance Dij(x, y) projects the move distance of each pixel on 0 °, 45 °, 90 ° and 135 ° four direction, then can obtain each
Move distance of the pixel in four directions of motion, and then the movement velocity of 0 °, 45 °, 90 ° and 135 ° four direction can be obtained
It is followed successively by：
D0°(i, j)=pixel is in the component in x directions
D90°(i, j)=pixel is in the component in y directions
Step 3: characteristic pattern is calculated with notable figure
The impression of retina also has center-surrounding features (Center-Surround) in human visual system.Center-
Surrounding features refer to that vision system is most sensitive regarding signal to central area, and it is inhibited to regard signal to peripheral region.Vision attention
The center of model-periphery operation is to carry out interlayer phase reducing to central stratum and peripheral tier, that is, it is corresponding to calculate different resolution
The difference of large and small scale, the mathematic(al) representation of characteristic pattern E are：
E (c, s)=| E (c) Θ E (s) | (13)
Wherein, Θ represents that the interpolation of two different scale images is subtracted each other, i.e., first to peripheral tier into row interpolation, pixel number is increased
The difference operation of central stratum and peripheral tier is carried out after to corresponding central stratum pixel number, c represents center scale, behalf neighboring area
Scale, c ∈ { 2,3,4 }, i.e., 2,3,4 layers are central stratum, s=c+ δ, scale difference δ ∈ { 3,4 }, i.e., each central stratum corresponding week
Boundary layer adds 3 or 4 for the number of plies of central stratum.
It can be in the hope of the characteristic pattern of gray scale, color, direction and motion feature according to above formula.
Gray feature figure：
I (c, s)=| I (c) Θ I (s) | (14)
Color characteristic figure：RG (c, s)=| (R (c)-G (c)) Θ (G (s)-R (s)) |
BY (c, s)=| (B (c)-Y (c)) Θ (Y (s)-B (s)) | (15)
Direction character figure：
O (c, s, θ)=| O (c, θ) Θ O (s, θ) | (16)
Motion feature figure：Ds(c, s)=| Ds(c)ΘDs(s)|
D0°(c, s)=| D0°(c)ΘD0°(s)|
D45°(c, s)=| D45°(c)ΘD45°(s)| (17)
D90°(c, s)=| D90°(c)ΘD90°(s)|
D135°(c, s)=| D135°(c)ΘD135°(s)|
All kinds of characteristic patterns are merged using default normalization operator N (), obtain the notable figure of each category feature.
Gray scale notable figure：
Color notable figure：
Direction notable figure：
Movement velocity notable figure：
Direction of motion notable figure：
⊕ is default blending algorithm, and characteristic remarkable picture further normalizes, and addition obtains comprehensive notable figure：
After obtaining the synthesis notable figure of image, each position competes with one another in comprehensive notable figure, using " the victor is a king " mechanism,
The position of triumph becomes focus-of-attention, forms area-of-interest.The transfer of focus-of-attention is realized by " inhibition of return " mechanism.
The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, several improvement and deformation can also be made, these are improved and deformation
Also it should be regarded as protection scope of the present invention.
Claims (4)
1. a kind of region of interest area detecting method of view-based access control model attention model, which is characterized in that include the following steps：
Step 1: establish multi-scale image structure
One width two dimensional image, the scale space expression under different scale can be by input picture I (x, y) and Gaussian kernel G (x, y, σ)
Convolution obtains：
L (x, y, σ)=G (x, y, σ) * I (x, y) (2)
In formula (2), the location of pixels of (x, y) representative image, σ is known as the scale space factor, and value is smaller, characterizes the image quilt
Smooth is fewer, and corresponding scale is also just smaller, and large scale corresponds to the general picture feature of image, and small scale corresponds to the thin of image
Feature is saved, L represents the scale space of image；
Step 2: low-level visual features are extracted
1st, brightness extracts
If color video frequency image, r, g, b represent red, green and blue in image respectively, then brightness calculation formula is：
I=(r+g+b)/3 (3)
If gray level image, then gray feature directly chooses the gray value of each pixel；
2nd, color feature extracted
It is respectively red, green, blue, yellow 4 Color Channels to define R, G, B, Y, then：
Y=(r+g)/2- (r-g)/2-b
Then RG and BY channels are：RG=R-GB, Y=B- (R+G)/2；
3rd, Directional feature extraction
Direction character is extracted with Gabor filter：
One-dimensional Gabor functions, i.e. 1D-Gabor functions：
In formula, standard deviations of the σ for Gaussian function, w0For the spatial frequency of complex plane wave, x0Central point for 1D-Gabor functions is sat
Mark, the formula of the odd, even component of 1D-Gabor functions are respectively：
According to 3 σ principles, l=6 σ/w are taken0；
Two-dimensional Gabor function, i.e. 2D-Gabor functions are one under the Gauss envelopes multiple change sine waves along x-axis, two dimension
Gabor function expressions are：
The real and imaginary parts of two-dimensional Gabor function are respectively：
θ is directioin parameter, and σ is bigger, and energy more disperses, smaller more to concentrate, and has high frequency σh, low frequency σlTwo centre frequencies；
The selection of deflection θ：θ=0,30,60,90,120,150, filtering window 32*32 share 24 Gabor filtering
Device；
Input picture I (x, y) and the convolution of Gabor wavelet kernel function are：
R (x, y)=∫ ∫ I (ε, η) g (x- ε, y- η) d ε d η (8)
Gabor wavelet transformation after the result is that plural number, take plural number mould | | r (x, y) | | mean value and variance as wavelet transformation
As a result；
θ=0 is selected, the Gabor filter output of 30,60,90,120,150 six directions passes through (7) formula as direction character
The Gaobr wave filters of six direction are obtained, then each tomographic image of gaussian pyramid structure is filtered with these wave filters, is obtained
Feature Mapping figure on to corresponding six direction；
4th, Motion feature extraction
Motion feature includes movement velocity and two category feature of the direction of motion；
1) motion vector extracts：The motion vector of target is extracted using background subtraction, by calculating in image between each pixel
Time domain or spatial domain on relevance, using the background of former frame as reference, the background of present frame is calculated, then by current frame image
With its background subtracting, difference image is obtained, if t frames image is I (x, y, t), corresponding background image is B (x, y, t), then poor
Partial image is：
2) Motion feature extraction：According to the difference image, the move distance D of each pixel in adjacent two field pictures is obtainedij
(x, y) projects the move distance of each pixel on 0 °, 45 °, 90 ° and 135 ° four direction, then obtains each pixel in four fortune
Move distance on dynamic direction, and then the movement velocity for obtaining 0 °, 45 °, 90 ° and 135 ° four direction is followed successively by：
Step 3: characteristic pattern is calculated with notable figure
The center of visual attention model-periphery operation is to carry out interlayer phase reducing to central stratum and peripheral tier, that is, is calculated different
The difference of the corresponding large and small scale of resolution ratio, the mathematic(al) representation of characteristic pattern E are：
E (c, s)=| E (c) Θ E (s) | (11)
Wherein, Θ represents that the interpolation of two different scale images is subtracted each other, i.e., first to peripheral tier into row interpolation, pixel number is increased to pair
Carrying out the difference operation of central stratum and peripheral tier after the central stratum pixel number answered, c represents center scale, behalf neighboring area scale,
C ∈ { 2,3,4 }, i.e., 2,3,4 layers are central stratum, s=c+ δ, scale difference δ ∈ { 3,4 }, i.e., the corresponding peripheral tier of each central stratum is
The number of plies of central stratum adds 3 or 4；
The characteristic pattern of gray scale, color, direction and motion feature is acquired according to formula (11)：
Gray feature figure：
I (c, s)=| I (c) Θ I (s) | (12)
Color characteristic figure：RG (c, s)=| (R (c)-G (c)) Θ (G (s)-R (s)) |
BY (c, s)=| (B (c)-Y (c)) Θ (Y (s)-B (s)) | (13)
Direction character figure：
O (c, s, θ)=| O (c, θ) Θ O (s, θ) | (14)
Motion feature figure：
All kinds of characteristic patterns are merged using default normalization operator N (), obtain the notable figure of each category feature.
Gray scale notable figure：
Color notable figure：
Direction notable figure：
Movement velocity notable figure：
Direction of motion notable figure：
To preset blending algorithm, characteristic remarkable picture further normalizes, and addition obtains comprehensive notable figure：
After obtaining the synthesis notable figure of image, each position competes with one another in comprehensive notable figure, and the position of triumph becomes focus-of-attention,
Form area-of-interest.
2. the method as described in claim 1, which is characterized in that in step 3, comprehensive notable figure is inputted into Itti vision modes
In in the victor is a king network, select the region of conspicuousness maximum in notable figure successively as area-of-interest.
3. the method as described in claim 1, which is characterized in that further included after step 3 through returning in Itti vision modes
Return the step of suppression mechanism realizes the transfer of focus-of-attention.
4. the method as described in claims 1 or 2 or 3, which is characterized in that w0=1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810034712.3A CN108154147A (en) | 2018-01-15 | 2018-01-15 | The region of interest area detecting method of view-based access control model attention model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810034712.3A CN108154147A (en) | 2018-01-15 | 2018-01-15 | The region of interest area detecting method of view-based access control model attention model |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108154147A true CN108154147A (en) | 2018-06-12 |
Family
ID=62461249
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810034712.3A Pending CN108154147A (en) | 2018-01-15 | 2018-01-15 | The region of interest area detecting method of view-based access control model attention model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108154147A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109284670A (en) * | 2018-08-01 | 2019-01-29 | 清华大学 | A kind of pedestrian detection method and device based on multiple dimensioned attention mechanism |
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN110688783A (en) * | 2019-11-07 | 2020-01-14 | 济南大学 | Downhole pump target detection method based on MATLAB GUI |
| CN110956086A (en) * | 2019-10-24 | 2020-04-03 | 国家电网有限公司 | Machine room equipment online identification method based on image identification technology |
| CN111654647A (en) * | 2020-06-02 | 2020-09-11 | 苏州睿新微系统技术有限公司 | Induction imaging control method, device and control circuit of infrared detector |
| CN112699832A (en) * | 2021-01-12 | 2021-04-23 | 腾讯科技（深圳）有限公司 | Target detection method, device, equipment and storage medium |
| CN113780078A (en) * | 2021-08-05 | 2021-12-10 | 广州西威科智能科技有限公司 | Method for quickly and accurately identifying fault object in unmanned visual navigation |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100189354A1 (en) * | 2009-01-28 | 2010-07-29 | Xerox Corporation | Modeling images as sets of weighted features |
| CN103106672A (en) * | 2013-01-25 | 2013-05-15 | 西北工业大学 | Method for detecting interested region of image based on color features |
| CN104835175A (en) * | 2015-05-26 | 2015-08-12 | 西南科技大学 | Visual attention mechanism-based method for detecting target in nuclear environment |
| CN105825238A (en) * | 2016-03-30 | 2016-08-03 | 江苏大学 | Visual saliency object detection method |
| CN106462771A (en) * | 2016-08-05 | 2017-02-22 | 深圳大学 | 3D image significance detection method |
-
2018
- 2018-01-15 CN CN201810034712.3A patent/CN108154147A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20100189354A1 (en) * | 2009-01-28 | 2010-07-29 | Xerox Corporation | Modeling images as sets of weighted features |
| CN103106672A (en) * | 2013-01-25 | 2013-05-15 | 西北工业大学 | Method for detecting interested region of image based on color features |
| CN104835175A (en) * | 2015-05-26 | 2015-08-12 | 西南科技大学 | Visual attention mechanism-based method for detecting target in nuclear environment |
| CN105825238A (en) * | 2016-03-30 | 2016-08-03 | 江苏大学 | Visual saliency object detection method |
| CN106462771A (en) * | 2016-08-05 | 2017-02-22 | 深圳大学 | 3D image significance detection method |
Non-Patent Citations (4)
| Title |
|---|
| 均陈晓: "基于尺度不变特征变换的图像匹配1", 《百度文库》 * |
| 徐振辉: "一种基于改进视觉注意模型和局部自相似性的目标自动检测算法研究", 《科学技术与工程》 * |
| 曾姝彦等: "基于Gabor滤波器的图像目标识别方法", 《北京航天航空大学学报》 * |
| 韩成美: "图像感兴趣区域数字水印算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109284670A (en) * | 2018-08-01 | 2019-01-29 | 清华大学 | A kind of pedestrian detection method and device based on multiple dimensioned attention mechanism |
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN110956086A (en) * | 2019-10-24 | 2020-04-03 | 国家电网有限公司 | Machine room equipment online identification method based on image identification technology |
| CN110688783A (en) * | 2019-11-07 | 2020-01-14 | 济南大学 | Downhole pump target detection method based on MATLAB GUI |
| CN111654647A (en) * | 2020-06-02 | 2020-09-11 | 苏州睿新微系统技术有限公司 | Induction imaging control method, device and control circuit of infrared detector |
| CN111654647B (en) * | 2020-06-02 | 2022-12-09 | 苏州睿新微系统技术有限公司 | Induction imaging control method, device and control circuit of infrared detector |
| CN112699832A (en) * | 2021-01-12 | 2021-04-23 | 腾讯科技（深圳）有限公司 | Target detection method, device, equipment and storage medium |
| CN112699832B (en) * | 2021-01-12 | 2023-07-04 | 腾讯科技（深圳）有限公司 | Target detection method, device, equipment and storage medium |
| CN113780078A (en) * | 2021-08-05 | 2021-12-10 | 广州西威科智能科技有限公司 | Method for quickly and accurately identifying fault object in unmanned visual navigation |
| CN113780078B (en) * | 2021-08-05 | 2024-03-19 | 广州西威科智能科技有限公司 | Rapid and accurate fault object identification method in unmanned visual navigation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108154147A (en) | The region of interest area detecting method of view-based access control model attention model | |
| Zhang et al. | A categorization of multiscale-decomposition-based image fusion schemes with a performance study for a digital camera application | |
| Vincent et al. | A descriptive algorithm for sobel image edge detection | |
| Batool et al. | Fast detection of facial wrinkles based on Gabor features using image morphology and geometric constraints | |
| CN105550685B (en) | The large format remote sensing image area-of-interest exacting method of view-based access control model attention mechanism | |
| CN104463877B (en) | A kind of water front method for registering based on radar image Yu electronic chart information | |
| CN107066916B (en) | Scene semantic segmentation method based on deconvolution neural network | |
| WO2018145470A1 (en) | Image detection method and device | |
| CN104835175B (en) | Object detection method in a kind of nuclear environment of view-based access control model attention mechanism | |
| CN104217221A (en) | Method for detecting calligraphy and paintings based on textural features | |
| CN107330371A (en) | Acquisition methods, device and the storage device of the countenance of 3D facial models | |
| Draper et al. | Evaluation of selective attention under similarity transformations | |
| CN102844766A (en) | Human eyes images based multi-feature fusion identification method | |
| CN110163119A (en) | A kind of finger vein identification method and system | |
| CN105809173B (en) | A kind of image RSTN invariable attribute feature extraction and recognition methods based on bionical object visual transform | |
| CN101615245A (en) | Expression recognition method based on AVR and enhancing LBP | |
| CN107292907A (en) | A kind of method to following target to be positioned and follow equipment | |
| CN105023253A (en) | Visual underlying feature-based image enhancement method | |
| CN107945111A (en) | A kind of image split-joint method based on SURF feature extraction combination CS LBP descriptors | |
| CN110263605A (en) | Pedestrian's dress ornament color identification method and device based on two-dimension human body guise estimation | |
| CN115100719A (en) | Face recognition method based on fusion of Gabor binary pattern and three-dimensional gradient histogram features | |
| CN106599891A (en) | Remote sensing image region-of-interest rapid extraction method based on scale phase spectrum saliency | |
| Yang et al. | Adversarial reconstruction CNN for illumination-robust frontal face image recovery and recognition | |
| CN106296632B (en) | A kind of well-marked target detection method based on amplitude spectrum analysis | |
| Roy et al. | LGVTON: a landmark guided approach for model to person virtual try-on |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180612 |