CN103996040A - Bottom-up visual saliency generating method fusing local-global contrast ratio - Google Patents
Bottom-up visual saliency generating method fusing local-global contrast ratio Download PDFInfo
- Publication number
- CN103996040A CN103996040A CN201410200489.7A CN201410200489A CN103996040A CN 103996040 A CN103996040 A CN 103996040A CN 201410200489 A CN201410200489 A CN 201410200489A CN 103996040 A CN103996040 A CN 103996040A
- Authority
- CN
- China
- Prior art keywords
- segment
- image
- local
- size
- contrast
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a bottom-up visual saliency generating method fusing the local-global contrast ratio. According to the method, firstly, the local contrast ratio between a certain image block in an image and the other image blocks in a neighbor domain and the global contrast ratio between the image block and the remaining image blocks in the image are calculated based on the sparse coding theory, then, the two kinds of comparative information are organically combined together, the center offset is added into the two kinds of comparative information, finally, fusion of the local contrast ratio and the global contrast ratio is achieved, and a visual saliency calculation model with better accuracy and visual is built.
Description
Technical field
The invention belongs to computer vision algorithms make research field, relate to a kind of bottom-up vision significance generation method that merges part-global contrast, can be in natural image database accurately, calculate the remarkable figure of Given Graph picture robust.
Background technology
Vision significance is a critical function of vision attention, and it shows as observer and from a complicated visual scene, selects an important content to focus on, and ignores other not too important contents.In visual scene, some content more can obtain observer's attention than other guide, and we claim them to have higher vision significance.The thought of vision significance is applied in the computation model of vision attention in large quantities, and the significance measure method that ITTI adopts in its classical visual attention computation model is based on pixel and the local visual feature difference of neighborhood around it; The people such as Ma proposed a kind of significance measure method based on Characteristic Contrast in 2003, and the method is only considered color characteristic, is LUV space by input picture from RGB color space conversion, carries out color quantizing.Simple in order to process, input picture is adjusted to a fixing size.Calculating pixel and its be the color characteristic contrast of neighborhood around, obtains the conspicuousness value of this pixel; The people such as Hou have proposed a kind of significance measure method based on spectrum residual error in 2008, the method is analyzed the feature of marking area on frequency domain, build significantly figure in spatial domain.The people such as FengLiu proposed the significance measure based on region afterwards, and first they utilize certain method to obtain regions different in image, then according to the positional factor in each region and its conspicuousness of Characteristic Contrast isometry.
Although above-mentioned conspicuousness computation model can draw gratifying result in specific Sample Storehouse, but in these models, still there is an obvious defect: they all only considered in the global contrast of image or local contrast a bit, and the information architecture of simultaneously not applying these two kinds contrast property goes out a unified conspicuousness computation model.Experiment shows, the marking area based on local feature contrast easily concentrates on the background area of the stronger marginal portion of variation or more complicated; The marking area of the conspicuousness based on global characteristics contrast can not finely highlight and the region that has sharp contrast degree around.Based on this, the present invention proposes a kind of bottom-up vision significance computing method that merge part-global contrast, first according to theoretical local contrast and the global contrast feature of extracting in image of sparse coding, then the information of these two kinds contrast property is organically combined, again according to the center-biased theory of human visual attention psychological study, set up out one and have more accuracy, the vision significance computation model of robustness.
Summary of the invention
The technical matters solving
For fear of the deficiencies in the prior art part, the present invention proposes a kind of bottom-up vision significance generation method that merges part-global contrast.
Technical scheme
Bottom-up vision significance computing method that merge part-global contrast, is characterized in that step is as follows:
Step 1 is extracted segment and the feature thereof in image: be first N × N pixel by image down sampling, then adopting size is size ∈ [5,50], and step-length is
the image of square moving window after down-sampling in extract segment p
i, segment p
ithe vector that interior pixel value forms is using the feature x as this segment
i; Wherein i ∈ [1, M], M is the segment number in piece image;
Step 2 builds the local dictionary of segment pi: adopt size for size ∈ [5,50], step-length is
square moving window at segment p
iface in territory and to extract all and p
ioverlapping area is less than
segment, the matrix that the feature of these segments is formed is as segment p
ilocal dictionary
wherein segment p
iface territory size for Sru
size=β size, β ∈ [3,9] is the scale-up factor that faces territory scope;
Step 3 is calculated segment p
ilocal contrast: according to sparse coding theory, adopt segment p
ilocal dictionary
to its feature x
iencode:
wherein
the local sparse coding of current segment,
the local residual error after sparse coding, segment p
ilocal contrast
Step 4 builds segment p
iglobal Dictionary; Adopting size is size ∈ [5,50], and step-length is
the picture in its entirety of square moving window after down-sampling within the scope of extract all and segment p
ioverlapping area is less than
segment, the matrix that the feature of these segments is formed is as segment p
iglobal Dictionary
Step 5 is calculated segment p
iglobal contrast: according to sparse coding theory, adopt segment p
iglobal Dictionary
to its feature x
iencode:
wherein
the overall sparse coding of current segment,
the overall residual error after sparse coding, segment p
iglobal contrast
Step 6 is calculated segment p
icenter offset: calculate segment p
icenter offset
wherein: D
maxfor the distance farthest of the image middle distance image center after down-sampling; D
ifor segment p
ithe distance of the image middle distance image center of central point after down-sampling;
Step 7 is calculated segment p
iremarkable value: to segment p
ilocal contrast and global contrast merge
draw its significantly value S, wherein λ ∈ [01] is the weight coefficient of local contrast and global contrast;
Step 8 generates significantly figure: the remarkable value of calculating all segments in the image after down-sampling according to step 1-7, using these, significantly value is as the gray-scale value generation gray-scale map corresponding with image after down-sampling of segment corresponding thereto, and the size that this gray-scale map is upsampled to original image is the remarkable figure of synthetic image;
In described step 3 and 5, the method for compute sparse coefficient and residual error adopts document Han B, Zhu H, Ding Y. " Bottom-up saliency based on weighted sparse coding residual ", Proceedings of the19th ACM international conference on Multimedia.ACM, the method for 2011:1117-1120.
Beneficial effect
The present invention proposes a kind of bottom-up vision significance computing method that merge part-global contrast, first utilize in the theoretical computed image of sparse coding in the local contrast between some image blocks and other image blocks within it faces territory and this image block and image and remain the global contrast between all image blocks, then the information of these two kinds contrast property is organically combined and adds center offset, finally realize local contrast, the fusion of global contrast, set up out one and have more accuracy, the vision significance computation model of robustness.
Attached caption
Fig. 1: the basic flow sheet of the inventive method
Fig. 2: experiment comparing result figure
Fig. 3: ROC result figure
Embodiment
Now in conjunction with the embodiments, the invention will be further described for accompanying drawing:
For the hardware environment of implementing be: Intel Pentium2.93GHz CPU computing machine, 2.0GB internal memory, the software environment of operation is: Matlab R2011b and Windows XP.Testing all images of having chosen in BRUCE storehouse as test data, comprise 120 width natural images in this database, is the database for testing vision conspicuousness computation model of International Publication.
The present invention is specifically implemented as follows:
1. extract segment and the feature thereof in image: be first N × N pixel by image down sampling, then adopting size is size ∈ [5,50], and step-length is
the image of square moving window after down-sampling in extract segment p
i, segment p
ithe vector that interior pixel value forms is using the feature x as this segment
i; Wherein i ∈ [1, M], M is the segment number in piece image.
2. build the local dictionary of segment pi: adopting size is size ∈ [5,50], and step-length is
square moving window at segment p
iface in territory and to extract all and p
ioverlapping area is less than
segment, the matrix that the feature of these segments is formed is as segment p
ilocal dictionary
wherein segment p
iface territory size for Sru
size=β size, β ∈ [3,9] is the scale-up factor that faces territory scope.
3. calculate the local contrast of segment pi: according to the method in sparse coding theory and " Bottom-up saliency based on weighted sparse coding residual ", adopt segment p
ilocal dictionary
to its feature x
iencode:
wherein
the local sparse coding of current segment,
the local residual error after sparse coding, segment p
ilocal contrast
4. build segment p
iglobal Dictionary; Adopting size is size ∈ [5,50], and step-length is
the picture in its entirety of square moving window after down-sampling within the scope of extract all and segment p
ioverlapping area is less than
segment, the matrix that the feature of these segments is formed is as segment p
iglobal Dictionary
5. calculate the global contrast of segment pi: according to the method in sparse coding theory and " Bottom-up saliency based on weighted sparse coding residual ", adopt segment p
iglobal Dictionary
to its feature x
iencode:
wherein
the overall sparse coding of current segment,
the overall residual error after sparse coding, segment p
iglobal contrast
6. calculate segment p
icenter offset: calculate segment p
icenter offset
wherein: D
maxfor the distance farthest of the image middle distance image center after down-sampling; D
ifor segment p
ithe distance of the image middle distance image center of central point after down-sampling.
7. calculate segment p
iremarkable value: to segment p
ilocal contrast and global contrast merge
draw its significantly value S
i, wherein λ ∈ [0,1] is the weight coefficient of local contrast and global contrast.
8. generate significantly figure: the remarkable value of calculating all segments in the image after down-sampling according to step 1-7, using these, significantly value is as the gray-scale value generation gray-scale map corresponding with image after down-sampling of segment corresponding thereto, and the size that this gray-scale map is upsampled to original image is the remarkable figure of synthetic image.
The present invention selects ROC curve to assess recognition result.This curve definitions is under segmentation threshold changes, the variation relation of false alarm rate (FPR) and recall rate (TPR).Computing formula is as follows:
Wherein FP is the false-alarm region detecting, N is non-order target area in ground truth; TP is the real police region territory detecting, P is order target area in ground truth.
Accompanying drawing 2 is some contrast and experiment, wherein, CS refers to the remarkable figure that the local contrast only utilized in the present invention calculates, CG refers to the remarkable figure that the global contrast only utilized in the present invention calculates, and CS+CG is the remarkable figure calculating according to the method that merges part-global contrast in the present invention.Can find out that algorithm that the present invention proposes can overcome the defect that independent use local contrast or global contrast are brought, can be in natural image database accurately, calculate the remarkable figure of Given Graph picture robust.The ROC curve that accompanying drawing 3 is the inventive method, table 1 is the quantitative contrast result of the inventive method and other existing algorithms, in table, the value of secondary series is respective algorithms ROC area under a curve (AUC) in BRUCE test library, from experimental result can find out method that the present invention proposes can be more accurately and robust natural image is carried out to the calculating of remarkable figure.
The contrast of table 1 conspicuousness testing result
Methods: | AIM | Itti’s | Judd’s | Liyin’s | Hanbiao’s | OURS |
AUC: | 0.7241 | 0.7455 | 0.7795 | 0.8006 | 0.8264 | 0.8360 |
Claims (2)
1. one kind merges the bottom-up vision significance generation method of part-global contrast, it is characterized in that step is as follows:
Step 1 is extracted segment and the feature thereof in image: be first N × N pixel by image down sampling, then adopting size is size ∈ [5,50], and step-length is
the image of square moving window after down-sampling in extract segment p
i, segment p
ithe vector that interior pixel value forms is using the feature x as this segment
i; Wherein i ∈ [1, M], M is the segment number in piece image;
Step 2 builds the local dictionary of segment pi: adopt size for size ∈ [5,50], step-length is
square moving window at segment p
iface in territory and to extract all and p
ioverlapping area is less than
segment, the matrix that the feature of these segments is formed is as segment p
ilocal dictionary
wherein segment p
iface territory size for Sru
size=β size, β ∈ [3,9] is the scale-up factor that faces territory scope;
Step 3 is calculated segment p
ilocal contrast: according to sparse coding theory, adopt segment p
ilocal dictionary
to its feature x
iencode:
wherein
the local sparse coding of current segment,
the local residual error after sparse coding, segment p
ilocal contrast
Step 4 builds segment p
iglobal Dictionary; Adopting size is size ∈ [5,50], and step-length is
the picture in its entirety of square moving window after down-sampling within the scope of extract all and segment p
ioverlapping area is less than
the segment of e, the matrix that the feature of these segments is formed is as segment p
iglobal Dictionary
Step 5 is calculated segment p
iglobal contrast: according to sparse coding theory, adopt segment p
iglobal Dictionary
to its feature x
iencode:
wherein
the overall sparse coding of current segment,
the overall residual error after sparse coding, segment p
iglobal contrast
Step 6 is calculated segment p
icenter offset: calculate segment p
icenter offset
wherein: D
maxfor the distance farthest of the image middle distance image center after down-sampling; D
ifor segment p
ithe distance of the image middle distance image center of central point after down-sampling;
Step 7 is calculated segment p
iremarkable value: to segment p
ilocal contrast and global contrast merge
) go out its significantly value S
i, wherein λ ∈ [0,1] is the weight coefficient of local contrast and global contrast;
Step 8 generates significantly figure: the remarkable value of calculating all segments in the image after down-sampling according to step 1-7, using these, significantly value is as the gray-scale value generation gray-scale map corresponding with image after down-sampling of segment corresponding thereto, and the size that this gray-scale map is upsampled to original image is the remarkable figure of synthetic image.
2. the bottom-up vision significance generation method of fusion part-global contrast according to claim 1, it is characterized in that: in described step 3 and 5, the method for compute sparse coefficient and residual error adopts document Han B, Zhu H, Ding Y. " Bottom-up saliency based on weighted sparse coding residual ", Proceedings of the19th ACM international conference on Multimedia.ACM, the method for 2011:1117-1120.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410200489.7A CN103996040A (en) | 2014-05-13 | 2014-05-13 | Bottom-up visual saliency generating method fusing local-global contrast ratio |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201410200489.7A CN103996040A (en) | 2014-05-13 | 2014-05-13 | Bottom-up visual saliency generating method fusing local-global contrast ratio |
Publications (1)
Publication Number | Publication Date |
---|---|
CN103996040A true CN103996040A (en) | 2014-08-20 |
Family
ID=51310201
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201410200489.7A Pending CN103996040A (en) | 2014-05-13 | 2014-05-13 | Bottom-up visual saliency generating method fusing local-global contrast ratio |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN103996040A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104408708A (en) * | 2014-10-29 | 2015-03-11 | 兰州理工大学 | Global-local-low-rank-based image salient target detection method |
CN106295542A (en) * | 2016-08-03 | 2017-01-04 | 江苏大学 | A kind of road target extracting method of based on significance in night vision infrared image |
CN106709512A (en) * | 2016-12-09 | 2017-05-24 | 河海大学 | Infrared target detection method based on local sparse representation and contrast |
CN107423765A (en) * | 2017-07-28 | 2017-12-01 | 福州大学 | Based on sparse coding feedback network from the upper well-marked target detection method in bottom |
CN107886533A (en) * | 2017-10-26 | 2018-04-06 | 深圳大学 | Vision significance detection method, device, equipment and the storage medium of stereo-picture |
CN110245660A (en) * | 2019-06-03 | 2019-09-17 | 西北工业大学 | Webpage based on significant characteristics fusion sweeps path prediction technique |
CN114494262A (en) * | 2022-04-19 | 2022-05-13 | 广东粤港澳大湾区硬科技创新研究院 | Method and device for evaluating image contrast |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102831402A (en) * | 2012-08-09 | 2012-12-19 | 西北工业大学 | Sparse coding and visual saliency-based method for detecting airport through infrared remote sensing image |
US20140126824A1 (en) * | 2012-11-05 | 2014-05-08 | Raytheon Bbn Technologies Corp. | Efficient inner product computation for image and video analysis |
-
2014
- 2014-05-13 CN CN201410200489.7A patent/CN103996040A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102831402A (en) * | 2012-08-09 | 2012-12-19 | 西北工业大学 | Sparse coding and visual saliency-based method for detecting airport through infrared remote sensing image |
US20140126824A1 (en) * | 2012-11-05 | 2014-05-08 | Raytheon Bbn Technologies Corp. | Efficient inner product computation for image and video analysis |
Non-Patent Citations (3)
Title |
---|
HAN B等: "bottom-up saliency based on weighted sparse coding residual", 《PROCEEDINGS OF THE19TH ACM INTERNATIONAL CONFERENCE ON MULTIMEDIA.ACM》 * |
张杰: "自底向上的视觉显著区域自动提取技术研究", 《中国优秀硕士学位论文全文数据库(电子期刊) 信息科技辑》 * |
杨维斌: "自底向上的视觉显著性检测方法与应用研究", 《万方数据库》 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104408708A (en) * | 2014-10-29 | 2015-03-11 | 兰州理工大学 | Global-local-low-rank-based image salient target detection method |
CN104408708B (en) * | 2014-10-29 | 2017-06-20 | 兰州理工大学 | A kind of image well-marked target detection method based on global and local low-rank |
CN106295542A (en) * | 2016-08-03 | 2017-01-04 | 江苏大学 | A kind of road target extracting method of based on significance in night vision infrared image |
CN106709512A (en) * | 2016-12-09 | 2017-05-24 | 河海大学 | Infrared target detection method based on local sparse representation and contrast |
CN107423765A (en) * | 2017-07-28 | 2017-12-01 | 福州大学 | Based on sparse coding feedback network from the upper well-marked target detection method in bottom |
CN107886533A (en) * | 2017-10-26 | 2018-04-06 | 深圳大学 | Vision significance detection method, device, equipment and the storage medium of stereo-picture |
CN107886533B (en) * | 2017-10-26 | 2021-05-04 | 深圳大学 | Method, device and equipment for detecting visual saliency of three-dimensional image and storage medium |
CN110245660A (en) * | 2019-06-03 | 2019-09-17 | 西北工业大学 | Webpage based on significant characteristics fusion sweeps path prediction technique |
CN110245660B (en) * | 2019-06-03 | 2022-04-22 | 西北工业大学 | Webpage glance path prediction method based on saliency feature fusion |
CN114494262A (en) * | 2022-04-19 | 2022-05-13 | 广东粤港澳大湾区硬科技创新研究院 | Method and device for evaluating image contrast |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103996040A (en) | Bottom-up visual saliency generating method fusing local-global contrast ratio | |
Liu et al. | Automatic building extraction on high-resolution remote sensing imagery using deep convolutional encoder-decoder with spatial pyramid pooling | |
US10692218B2 (en) | Method and system of detecting image tampering, electronic device and storage medium | |
Mnih et al. | Learning to label aerial images from noisy data | |
Zhang et al. | A Linear Dirichlet Mixture Model for decomposing scenes: Application to analyzing urban functional zonings | |
Zhang et al. | Integrating bottom-up classification and top-down feedback for improving urban land-cover and functional-zone mapping | |
Soheily-Khah et al. | Generalized k-means-based clustering for temporal data under weighted and kernel time warp | |
CN104599292B (en) | A kind of anti-noise moving object detection algorithm decomposed based on low-rank matrix | |
CN102542295B (en) | Method for detecting landslip from remotely sensed image by adopting image classification technology | |
WO2018023734A1 (en) | Significance testing method for 3d image | |
CN113255915B (en) | Knowledge distillation method, device, equipment and medium based on structured instance graph | |
CN105825200A (en) | High-spectrum abnormal object detection method based on background dictionary learning and structure sparse expression | |
CN102422324B (en) | Age estimation device and method | |
CN102663754B (en) | Image matching calculation method based on regional Gaussian weighting | |
CN103955926A (en) | Method for remote sensing image change detection based on Semi-NMF | |
CN102982544B (en) | Many foreground object image interactive segmentation method | |
CN104463248A (en) | High-resolution remote sensing image airplane detecting method based on high-level feature extraction of depth boltzmann machine | |
CN102831598A (en) | Remote sensing image change detecting method with combination of multi-resolution NMF (non-negative matrix factorization) and Treelet | |
CN103745233B (en) | The hyperspectral image classification method migrated based on spatial information | |
CN105809182B (en) | Image classification method and device | |
CN103839257A (en) | Method for detecting changes of SAR images of generalized Gaussian K&I | |
CN104700412A (en) | Calculating method of visual salience drawing | |
CN104077609A (en) | Saliency detection method based on conditional random field | |
Su et al. | Dynamic classifier selection using spectral-spatial information for hyperspectral image classification | |
CN102663776B (en) | Violent movement detection method based on characteristic point analysis and device thereof |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20140820 |