CN104299230B - Image significance detection method utilizing red-black wavelet transform - Google Patents
Image significance detection method utilizing red-black wavelet transform Download PDFInfo
- Publication number
- CN104299230B CN104299230B CN201410497460.XA CN201410497460A CN104299230B CN 104299230 B CN104299230 B CN 104299230B CN 201410497460 A CN201410497460 A CN 201410497460A CN 104299230 B CN104299230 B CN 104299230B
- Authority
- CN
- China
- Prior art keywords
- red
- black
- image
- conversation
- small wave
- 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.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20048—Transform domain processing
- G06T2207/20064—Wavelet transform [DWT]
Abstract
The invention relates to an image significance detection method utilizing red-black wavelet transform, and belongs to the technical field of image analysis and processing. The method comprises the following steps that according to the size of an image I, the number of decomposition layers of red-black wavelets is determined, and the image I is subjected to red-black wavelet transform to obtain T; the image I is subjected to Gaussian smoothing processing and then red-black wavelet transform to obtain Tsmooth; the difference r between two results of red-black wavelet transform is calculated, r= Tsmooth-T, and r is subjected to red-black wavelet transform to obtain a significance image s; s is subjected to binary segmentation to obtain a significance detection result. Compared with a traditional frequency-domain analysis method, the method effectively improves accuracy of image significance detection.
Description
Technical field
The present invention relates to a kind of image significance detection method, more particularly, to a kind of image of utilization red and black small wave conversation shows
Work property detection method, belongs to image analyzing and processing technology field.
Background technology
In image domains, significance refers to that the mankind pass through visually-perceptible, captures the region attracting people's attention most in scene.
The significance detection of image is intended to the image according to input, finds out wherein significant region (prospect), it is opened with background segment
Come or be marked.Significance detection can be that other image related works provide and support, such as image segmentation, target recognition,
Image tagged, image adaptive compression etc..
At present, the most widely used technology solving saliency test problems is using bottom-up method, its
Basic thought assumes that other provincial characteristicss in the feature (as color, intensity, locus etc.) of prospect and entire image not
With there is uniqueness, therefore try to achieve and there is the characteristic area of uniqueness can get significance testing result.Using bottom-up
The usual step that method carries out significance detection is：
Step one, image segmentation is become with the close multiple regions of size, feature extraction is carried out to each region；
Step 2, by the feature between comparison domain, obtain the unique values in each region；
Step 3, it is calculated the significance value of each pixel according to the unique values in region；
Step 4, binarization segmentation is carried out to Saliency maps.
But the effect of significance detection often can be subject to mottled in a jumble texture in image, illumination jumpy, target
Relative scalar do not know etc. factor impact.
Compared to additive method, frequency domain analysises are commonly designed simply, calculate quick.Frequency-domain analysis method is a kind of pervasive
Detection method it is not necessary to building learning system and providing priori it is not necessary to various characteristics of image, such as color, texture
Consider Deng informix.It is its high efficiency calculating in place of frequency analysis method most worthy, be the real-time significance inspection of video
Survey provides probability.Image is carried out Fourier transformation by current frequency-domain analysis method, tries to achieve the residual error on frequency domain, then carries out
Fourier inversion to spatial domain obtains Saliency maps, and this method by frequency transformation extraction Saliency maps substantially increases
Calculating speed.
Although frequency domain analysises have above advantage, yet suffer from shortcoming：One side Saliency maps yardstick is more original
Picture size is much smaller, adds Gaussian Blur, the Saliency maps that the method obtains in another aspect Saliency maps calculating process
Foreground edge is fuzzyyer, and the foreground segmentation result obtaining is not fine.
Content of the invention
Present invention aim to address the Saliency maps fineness in traditional frequency-domain analysis method is relatively low it is difficult to acquisition is good
The problem of good salient region segmentation result, proposes a kind of image significance detection method of utilization red and black small wave conversation.
For achieving the above object, the technical solution adopted in the present invention is as follows：
A kind of image significance detection method of utilization red and black small wave conversation, comprises the following steps：
Step one, determine conversion number of plies L of red and black small wave conversation according to the size of input picture I by formula (1), according to L
I is carried out red and black small wave conversation and obtains T, T=BfwI, wherein BfwRepresent red and black small wave conversation；
L=[log2Min (w, h)] floor-1； (1)
Wherein, w and h represents width and the length of image I, [] respectivelyfloorRepresent and take downwards immediate integer value；
Step 2, original image I is carried out with Gaussian smoothing obtain Ismooth, according to L to IsmoothCarry out red-black small echo
Conversion obtains Tsmooth, Tsmooth=BfwIsmooth；
Step 3, try to achieve step one and difference r of step 2 transformation results, i.e. r=Tsmooth- T, and r is carried out red-black little
Ripple inverse transformation obtains Saliency maps s,WhereinRepresent red-black inverse wavelet transform,Represent to red and black
The element value of inverse wavelet transform result takes absolute value；
Step 4, according to formula (2) binarization segmentation carried out to Saliency maps s, obtain significance testing result；
Wherein, R (i, j) and s (i, j) represents at abscissa i, vertical coordinate j the binarization segmentation result of pixel and strong respectively
Degree, threshold value D=E (s), E (s) is the mathematic expectaion of each pixel intensity of Saliency maps.
Beneficial effect
Contrast traditional frequency domain analysis method, the inventive method has advantages below：
1st, Saliency maps fineness is high；
2nd, improve the accuracy of saliency detection；
3rd, the process time that can reach the detection of real-time saliency requires.
Brief description
Fig. 1 is the flow chart of the inventive method；
Fig. 2 is red and black small wave conversation schematic diagram in the principle of the invention；
Fig. 3 is red and black small wave conversation sub-band structure figure in the principle of the invention；
Fig. 4 is the red and black small wave conversation result being obtained by real image in the principle of the invention；
Fig. 5 is the Saliency maps being obtained by various sizes of wave filter in the embodiment of the present invention；
Fig. 6 is the inspection to one group of picture using institute's extracting method of the present invention and traditional frequency domain method Hou in the embodiment of the present invention
Survey results contrast, a is input picture, and b is the Saliency maps that red-black wavelet method obtains, C Hou method is calculated notable
Property figure, d be handmarking salient region, e is the salient region binary map being partitioned into using red-black wavelet method, and f is
The salient region binary map being partitioned into using Hou method；
Fig. 7 is the inspection to two groups of pictures using institute's extracting method of the present invention and traditional frequency domain method Hou in the embodiment of the present invention
Survey results contrast, its picture arrangement order is identical with Fig. 6；
Fig. 8 is the inspection to two groups of pictures using institute's extracting method of the present invention and traditional frequency domain method Hou in the embodiment of the present invention
Survey results contrast, its picture arrangement order is identical with Fig. 6.
Specific embodiment
In order to better illustrate technical scheme, below in conjunction with the accompanying drawings, by an embodiment, the present invention is done
Further illustrate.
Before introducing embodiment, introduce principle and the red and black small wave conversation method of the present invention first.
The ultimate principle of the inventive method is as follows：Red-black wavelet method is applied to the process of 2D signal, by signal
Successively carry out the lifting of horizontal and diagonally opposed lifting, image can be transformed to the signal on different frequency range.
The lifting of horizontal：The red and black that red-black wavelet method divides the image into similar checker-wise first is alternate
Two subsets, such as Fig. 2, is red subset R respectivelyk(useRepresent) and black subset (being represented with " ■ "), k represents ought
Front lifting level.Next using red subset RkPrediction black subset value, that is, for each black bars, using it week
, as the predictive value of black region, predictor formula is as follows for the meansigma methodss of four red square enclosing：
Wherein x represents the current image array lifting level, the value of x (i, j) representing matrix the i-th row jth column element.So
Afterwards, the new value based on black subset, updates red subset R by preserving statistical averagekValue：
The lifting of diagonal：Next, red subset R by firm renewalkIt is decomposed into two subsets：Blue subset Bk
(useRepresent) and yellow subset Yk(useRepresent).Using the prediction similar with horizontal vertical direction and renewal side
Formula processes blue and yellow subset.Forecast period is predicted to the value of yellow square using the data of Blue Squares：
The more new stage updates the value of Blue Squares using diagonally opposed yellow square：
Next, extract subset blue in matrix x obtaining new matrix y, y (i/2, j/2) ← x (i, j), works as i
During mod 2=j mod 2=0；So far just complete the wavelet transformation on k level.Using y as x continue above-mentioned red-black with blue-
Yellow decomposition can get the wavelet transform result on k+1 level.
It is understood that red-black wavelet method, in red-black catabolic phase, the value of black box can be by adjacent red
The value of color grid is predicted, and the value of black box is deducted the meansigma methodss of four adjacent red panels about, can be by
The uncertain component of red panels is saved in black box.In prediction steps, the value of red panels is added phase about
The 1/2 of the meansigma methodss of adjacent four black box, uncertain for black box component can be retained in red panels.Cause
This, remain the higher-frequency part (uncertain component) of signal level/vertical direction in black box, and red panels are protected
Stay the low frequency part of signal level/vertical direction.The decomposition meaning in blue-yellow stage can be released by red-black catabolic phase.Warp
Cross blue-yellow to decompose, blue grid remains the low frequency component on diagonal, and yellow panels remain higher-frequency on diagonal and divide
Amount.
Through once red-black wavelet decomposition, it is obtained in that red and black small wave conversation sub-band structure figure, as shown in Figure 3.Wherein DL
Corresponding to diagonally adjacent low frequency part, DH is diagonally adjacent HFS, and HVL corresponds to horizontal/vertical side
Low frequency part upwards.DL part is carried out red-black wavelet decomposition again, image can be decomposed further according to frequency band height
Become different sub-band.The transformation results of the red-black small echo being obtained by real image are as shown in Figure 4.
Embodiment：Saliency detects.On ASD data set using and test this method, include 1000 pictures,
And the salient region figure of accordingly artificial demarcation.
The image I that step one, reading data are concentrated, the size according to image I determines the conversion number of plies of red-black small echo, will scheme
Obtain T as I carries out red and black small wave conversation.
In order to remove the impact to testing result for the noise, original image size is contracted to the 1/4 of original size, according to contracting
Size after little determines the number of plies of wavelet transformation.For example, the image of 300 × 400, is changed into 150 × 200 after reducing,
According to formula (1), can convert the number of plies is L=[log2min(150,200)]floor- 1=6.According to the conversion number of plies, image is carried out
Red and black small wave conversation.The present embodiment carries out wavelet transform process on tri- different passages of RGB respectively.
Step 2, carry out Gaussian smoothing to I, then carry out red and black small wave conversation obtaining Tsmooth.
Identical with step one, original image size is contracted to the 1/4 of original size.Using different size of Gaussian smoothing
Wave filter, can obtain the Saliency maps of different-effect, as shown in Figure 5.The present embodiment adopts the Gaussian filter pair of 9 × 9 sizes
Image I is smoothed.Then on tri- different passages of RGB, wavelet transform process is carried out respectively to the image after smoothing.
Step 3, try to achieve step one and difference r of step 2 red and black small wave conversation result, r=Tsmooth- T, and r is carried out
Red-black inverse wavelet transform obtains Saliency maps s.
Because step one and step 2 have all carried out red and black small wave conversation to image in RGB triple channel respectively.Therefore exist
Their difference is calculated on RGB triple channel, and red-black inverse wavelet transform is carried out respectively to the difference result on three passages.To three
On passage, the inverse transformation result of red-black small echo carries out synthesis, such as each pixel is taken with the maximum of three passages, takes average etc., this
Embodiment is that the inverse transformation result in triple channel is averaged, and obtains Saliency maps.The resolution of the Saliency maps due to obtaining
Rate, again smaller than original image I size, can be used for interpolation method (as linear interpolation, non-linear interpolation etc.) and enter row interpolation obtaining
To and original image I with size Saliency maps, the present embodiment using two-way interpolation method Saliency maps are entered row interpolation obtain and
Original image is with the Saliency maps s of size.
Step 4, Saliency maps s is carried out with binarization segmentation, obtain significance testing result.
According to formula (2), Saliency maps are carried out binarization segmentation.Being worth the place for 1 is the salient region detecting, value
Place for 0 is background area.
Method proposed by the invention, although the original size of Saliency maps is still less than original image, due to being not required to
Gaussian filtering is carried out to Saliency maps, therefore Saliency maps reduction to be become after original image size, still can obtain relatively
Clearly details, has obtained more preferable testing result after binarization segmentation, such as Fig. 6 to Fig. 8.From fig. 6, it can be seen that and
Traditional frequency domain method is compared, and the Saliency maps being obtained using red-black wavelet method are finer, such as the strut of vault and horizontal stroke
Bar is all high-visible in Saliency maps 6b.In Fig. 7 and Fig. 8, the Saliency maps of traditional frequency domain method, between its foreground and background
Excessively very fuzzy, have impact on binarization segmentation result, for example, for the image of pedestrian and clock and watch, traditional frequency domain method can only obtain
Obtain the fraction region of human body and dial plate, and red-black wavelet method can obtain finer Saliency maps, shows acquisition
Finer object edge, such as petal edge, animal edge, complete foreground zone can be partitioned into according to this Saliency maps
Domain.Further for the background that texture is mottled, there is a certain degree of anti-interference, the mottled, blue sky on such as meadow and ground are handed over
Meet place etc., do not have to affect the detection to prospect.
Table 1 is the two kinds of inspections calculating on all 1000 pictures in ASD data base to this method and traditional frequency domain method
The hit rate (Hit Rate, HR) of survey method and the average result of false alarm rate (False Alarm Rate, FAR).Permissible by table 1
Find out, compared to traditional frequency-domain analysis method, this method obtains higher hit rate (0.8252 to 0.2278) and lower
False alarm rate (0.0605 to 0.2258), improves the significance Detection results of frequency analysis method.
Table 1 significance detection method results contrast
The real-time of saliency detection is also an important measurement index of significance detection.Table 2 illustrates this
Method and the comparison of multiple significance detection method process time.Traditional frequency domain method and red-black wavelet method both frequency domains
Analysis method, to the process time of every two field picture less than 0.1 second, average each second can process the image about 20 frames, therefore
It is applied to the real-time processing of video.And region contrast method and histogram contrast's method, employ bottom-up algorithm
Design, from the color characteristic of image, region contrast angularly calculates Saliency maps, and process time is much larger than frequency-domain analysiss
Method (region contrast method about every five seconds for example processes a frame picture, and histogram contrast's method is per second to process a frame picture).
Table 2 time statistical result is compared
Present invention uses red-black wavelet method is it is achieved that detect to the significance of image, and pass through experiment show,
Compared to traditional frequency domain method, this algorithm not only maintains the real-time of the algorithm of frequency-domain analysis method, also significantly carries simultaneously
The high accuracy of saliency detection.
The foregoing is only the specific embodiment of the present invention, the protection domain being not intended to limit the present invention, all at this
Within bright spirit and principle, any modification, equivalent substitution and improvement done etc., should be included in protection scope of the present invention
Within.
Claims (6)
1. a kind of image significance detection method of utilization red and black small wave conversation it is characterised in that：Comprise the following steps：
Step one：Determine conversion number of plies L of red and black small wave conversation by formula (1) according to the size of input picture I, according to L, I is entered
Row red and black small wave conversation obtains T, T=BfwI, wherein BfwRepresent red and black small wave conversation；
L=[log2Min (w, h)]floor-1； (1)
Wherein, w and h represents width and the length of image I, [] respectivelyfloorRepresent and take downwards immediate integer value；
Step 2：Gaussian smoothing is carried out to image I and obtains Ismooth, according to L to IsmoothCarry out red and black small wave conversation to obtain
Tsmooth, Tsmooth=BfwIsmooth；
Step 3：Try to achieve TsmoothWith difference r of T, i.e. r=Tsmooth- T, and r is carried out with red-black inverse wavelet transform acquisition significance
Figure s,WhereinRepresent red-black inverse wavelet transform,Represent the unit to red-black inverse wavelet transform result
Plain value takes absolute value；
Step 4：According to formula (2), binarization segmentation is carried out to Saliency maps s, obtains significance testing result；
Wherein, R (i, j) and s (i, j) represents the binarization segmentation result of pixel and intensity, threshold at abscissa i, vertical coordinate j respectively
Value D=E (s), E (s) is the mathematic expectaion of each pixel intensity of Saliency maps.
2. a kind of utilization red and black small wave conversation according to claim 1 image significance detection method it is characterised in that：
The described red and black small wave conversation that carries out is to carry out wavelet transformation respectively on tri- different passages of RGB.
3. a kind of utilization red and black small wave conversation according to claim 1 image significance detection method it is characterised in that：
Described Gaussian smoothing that image I is carried out is using Gaussian filter, image I to be smoothed.
4. a kind of utilization red and black small wave conversation according to claim 1 image significance detection method it is characterised in that：
Described try to achieve TsmoothWith difference r of T, and r is carried out with red-black inverse wavelet transform is respectively calculating difference in RGB triple channel, and
Red-black inverse wavelet transform is carried out to difference, then the inverse transformation result in triple channel is carried out synthesis.
5. a kind of utilization red and black small wave conversation according to claim 4 image significance detection method it is characterised in that：
Described synthesis is to average.
6. a kind of utilization red and black small wave conversation according to claim 1 image significance detection method it is characterised in that：
To inverse transformation result using interpolation method enter row interpolation obtain and described image I with size Saliency maps s.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410497460.XA CN104299230B (en) | 2014-09-25 | 2014-09-25 | Image significance detection method utilizing red-black wavelet transform |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410497460.XA CN104299230B (en) | 2014-09-25 | 2014-09-25 | Image significance detection method utilizing red-black wavelet transform |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104299230A CN104299230A (en) | 2015-01-21 |
| CN104299230B true CN104299230B (en) | 2017-02-15 |
Family
ID=52318952
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201410497460.XA Expired - Fee Related CN104299230B (en) | 2014-09-25 | 2014-09-25 | Image significance detection method utilizing red-black wavelet transform |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104299230B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108109138B (en) * | 2017-12-15 | 2021-08-06 | 华南理工大学 | Method for self-adaptive light uniformization of high-light area of mirror-like object |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101847203A (en) * | 2009-03-25 | 2010-09-29 | 何玉青 | Human face recognition method based on red and black small wave conversation |
| CN103247059A (en) * | 2013-05-27 | 2013-08-14 | 北京师范大学 | Remote sensing image region of interest detection method based on integer wavelets and visual features |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8326046B2 (en) * | 2009-02-11 | 2012-12-04 | Ecole De Technologie Superieure | Method and system for determining structural similarity between images |
-
2014
- 2014-09-25 CN CN201410497460.XA patent/CN104299230B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101847203A (en) * | 2009-03-25 | 2010-09-29 | 何玉青 | Human face recognition method based on red and black small wave conversation |
| CN103247059A (en) * | 2013-05-27 | 2013-08-14 | 北京师范大学 | Remote sensing image region of interest detection method based on integer wavelets and visual features |
Non-Patent Citations (3)
| Title |
|---|
| 基于HSV和红黑小波变换的多光谱图像融合;祝青 等;《计算机工程》;20120229;第38卷(第3期);第287-289页 * |
| 基于Red-Black小波变换的多光谱图像融合方法;刘斌 等;《仪器仪表学报》;20110228;第32卷(第2期);第408-414页 * |
| 基于红-黑小波变换的图像数据压缩方法;刘家胜 等;《计算机工程》;20050430;第31卷(第7期);第180-182页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104299230A (en) | 2015-01-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111080620B (en) | Road disease detection method based on deep learning | |
| CN105046677B (en) | A kind of enhancing treating method and apparatus for traffic video image | |
| Chen et al. | An improved license plate location method based on edge detection | |
| Peng et al. | A complete passive blind image copy-move forensics scheme based on compound statistics features | |
| CN102609701B (en) | Remote sensing detection method based on optimal scale for high-resolution SAR (synthetic aperture radar) | |
| CN102800086B (en) | Offshore scene significance detection method | |
| CN104268847A (en) | Infrared light image and visible light image fusion method based on interactive non-local average filtering | |
| CN101527043B (en) | Video picture segmentation method based on moving target outline information | |
| CN105096280A (en) | Method and device for processing image noise | |
| CN102163284A (en) | Chinese environment-oriented complex scene text positioning method | |
| CN103208097B (en) | Filtering method is worked in coordination with in the principal component analysis of the multi-direction morphosis grouping of image | |
| CN104504652A (en) | Image denoising method capable of quickly and effectively retaining edge and directional characteristics | |
| CN108932699B (en) | Three-dimensional matching harmonic filtering image denoising method based on transform domain | |
| CN104424641A (en) | Detection method for image fuzzy tampering | |
| CN110570404A (en) | Sobel operator-based cloth defect detection method | |
| CN111091111A (en) | Vehicle bottom dangerous target identification method | |
| CN103337073B (en) | A kind of two dimensional image threshold segmentation method based on three-dimensional entropy | |
| Zhang et al. | Single image dehazing based on fast wavelet transform with weighted image fusion | |
| CN107818547A (en) | The minimizing technology of the spiced salt and Gaussian mixed noise in a kind of sequence towards twilight image | |
| CN104463822A (en) | Multi-focus image fusing method and device based on multi-scale overall filtering | |
| CN104299230B (en) | Image significance detection method utilizing red-black wavelet transform | |
| Ein-shoka et al. | Quality enhancement of infrared images using dynamic fuzzy histogram equalization and high pass adaptation in DWT | |
| CN103632373B (en) | A kind of flco detection method of three-frame difference high-order statistic combination OTSU algorithms | |
| Li et al. | A study of crack detection algorithm | |
| Wu et al. | Research on crack detection algorithm of asphalt pavement |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170215 Termination date: 20170925 |