CN109886267A - A kind of soft image conspicuousness detection method based on optimal feature selection - Google Patents
A kind of soft image conspicuousness detection method based on optimal feature selection Download PDFInfo
- Publication number
- CN109886267A CN109886267A CN201910086101.8A CN201910086101A CN109886267A CN 109886267 A CN109886267 A CN 109886267A CN 201910086101 A CN201910086101 A CN 201910086101A CN 109886267 A CN109886267 A CN 109886267A
- Authority
- CN
- China
- Prior art keywords
- image
- pixel
- super
- conspicuousness
- saliency maps
- 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 soft image conspicuousness detection method based on optimal feature selection that the invention discloses a kind of.The present invention hierarchically refines Saliency maps by the best features being recursively adaptive selected in soft image.Inhibit background interference first with multiple dimensioned super-pixel segmentation.Then, initial Saliency maps are generated by global contrast and spatial relationship.The final Saliency maps optimized using part and global adaptability.And test on four data sets using the method, experimental evaluation shows proposed model in terms of efficiency and accuracy better than 15 kinds of state-of-the-art methods.It include: MSRA data set, SOD data set, DUT-OMRON data set and NI data set.
Description
Technical field
The soft image conspicuousness detection method based on optimal feature selection that the present invention relates to a kind of, belongs at image
Reason field can provide theory and technology basis for hot issues such as night safe monitoring, the positioning of complex environment target.
Background technique
Saliency detection can greatly promote the extensive use that computer is independently detected and identified.Although scheming in recent years
It has been achieved for being markedly improved as conspicuousness detects, but low signal-to-noise ratio, the characteristic image of low contrast is still to current method
Propose huge challenge.This patent proposes a kind of significant seed dispersal method based on optimal feature selection to detect low comparison
Spend the significant object in image.The key idea of the method proposed is by the way that soft image is recursively adaptive selected
In optimal characteristics hierarchically refine Saliency maps.It is intended to imitate human visual system, it can be like a dream from scene
Maximally related object is picked out, obvious object detection can greatly promote image segmentation, image retrieval, compression of images, wireless network
The application of network node distribution etc..
By calculating the pixel in low level or high-level clue or the uniqueness in region, existing conspicuousness model can be with
It is divided into two classes.1) one kind is to be mainly based upon locally or globally contrast from the upper unsupervised method in bottom.These algorithms have
Center ring based on multiple characteristic patterns is calculated around difference, is had and is carried out conspicuousness calculating based on histogram and region contrast, has
Estimate saliency using contrast and spatial distribution, have through higher-dimension colour switching and regression estimates overall situation conspicuousness and
Local conspicuousness is had and is assumed by using color and the compactedness of textural characteristics to construct Saliency maps.In general, these bottom of from and
On method meet difficulty when handling the image of mixed and disorderly background, and be difficult to find that really when picture contrast is relatively low
Significant object.
2) it is another kind of be from top and under the method that Target Acquisition is instructed by supervised learning.By method using support
Vector machine (SVM) training generates super-pixel grade notable figure, and some is propagated and trained by merging the Laplce based on super-pixel
Convolutional neural networks (CNN) design a kind of conspicuousness model based on deep learning, some propositions combine part and complete
The simplification CNN of office's feature, also some is proposed one and is learnt saliency based on the CNN model of covariance and utilize edge
Retain with multiple dimensioned context neural network and generates notable figure.These methods have higher computation complexity, current depth
The performance that neural network model is very time-consuming and they are in terms of accurate positioning is relatively weak.With in recent years to deep learning
Research go deep into, using convolutional neural networks as the network model of Typical Representative because its powerful learning ability is widely paid close attention to
And it is successfully applied to different visual tasks.A kind of model of the convolutional neural networks as simulation human brain neuromechanism,
The Object identifying that can complete similar human perception performance also can be considered that a kind of advanced significant clue is applied to soft image
In significant object detection.
Although have been proposed it is many upper the bottom of from and from top and under conspicuousness model, most of which quilt
Designed for the obvious object detection in scene on daytime.Due to indicating the useful feature of conspicuousness information in soft image
Compare deficient, these models may face huge challenge in low light scene.In terms of its reason is mainly following two: 1) mesh
The feature of preceding hand-designed is difficult to assess the objectivity in image；2) current advanced features are detecting accurate significant object edges
Huge challenge is faced in terms of boundary, the multistage convolution sum pond layer in CNN model makes significant object bounds very fuzzy.
Summary of the invention
The soft image conspicuousness detection method based on optimal feature selection that the present invention relates to a kind of.Saliency
Detection can greatly promote the extensive use that computer is independently detected and identified.Although saliency detection in recent years has taken
Significant improvement, but low signal-to-noise ratio were obtained, the characteristic image of low contrast still proposes huge challenge to current method.This is specially
Benefit proposes a kind of significant seed dispersal method based on optimal feature selection to detect the significant object in soft image.
The key idea of the method proposed is the best features by being recursively adaptive selected in soft image to be layered
Ground refines Saliency maps.Inhibit background interference first with multiple dimensioned super-pixel segmentation.Then, pass through global contrast and sky
Between relationship generate initial Saliency maps.The final Saliency maps optimized using part and global adaptability.And utilize this
Method test on four data sets, and experimental evaluation shows that proposed model is better than in terms of efficiency and accuracy
15 kinds of state-of-the-art methods.The best features selection method that soft image conspicuousness reproduces, the technology specifically comprises the steps of:
Step (1) is to retain and utilize the information of image, divides the image into super-pixel；
Step (2) extracts several visual signatures and therefrom selects optimal feature；
Step (3) constructs conspicuousness model and generates initial notable figure；
Step (4) recursively refines initial conspicuousness seed and optimizes Saliency maps.
In step (1), for object of reservation structural information and using the intermediate information of original image, changed using simple linear
Input picture is divided into super-pixel by generation cluster (SLIC) algorithm, is expressed as { si, i=1, N.The processing can lead to
It crosses and super-pixel is considered as processing unit to improve the efficiency of model.Since the accuracy of testing result is highly dependent on super-pixel
Quantity, the model proposed capture the super-pixel of three kinds of different scales, and wherein N is respectively set to 100,200 and 300.
In step (2), extract several lower-level vision features from input picture, including color characteristic, textural characteristics,
Direction character and Gradient Features, wherein color characteristic has nine kinds；The validity of lower-level vision feature is according to the difference of input picture
Contrast and change, the therefrom adaptively selected coldest days of the year end optimal characteristics of the comentropy based on lower-level vision feature.Lower-level vision is special
Sign extraction process is described as follows:
2-1. is first normalized the RGB color of input picture, to eliminate the influence of light and shade.Then will
Image is converted to LAB, HSV and YCbCr color space after normalization, to extract nine kinds of color characteristics.
2-2. indicates its textural characteristics using the two-dimensional entropy of image.The variation of textural characteristics is by the variation of entropy come really
Fixed, textural characteristics have the ability of very strong antinoise and geometry deformation.
2-3. is filtered by executing the Gabor of different directions θ ∈ { 0 °, 45 °, 90 °, 135 ° } on the gray level image of input
Wave device obtains direction character.It is smaller that the global property and rotational invariance of direction character influence it by low contrast.
2-4. calculates Gradient Features by average level gradient and vertical gradient.Part can be described by Gradient Features
The amplitude of grey scale change.
After the 1 dimension entropy by calculating each characteristic pattern, from extract 12 features L, A, B, H, S, V, Y, Cb, Cr,
E, O, G } in select nine optimal characteristics, nine optimal characteristics are expressed as { Fk, k=1,2 ..., 9 is its method are as follows:
Here pIIndicate the ratio I of the pixel where gray value.As characteristic statistics form, the polymerization of image grayscale distribution
The average information for including in characteristic can be reflected by image entropy.The Image entropy of characteristic pattern is bigger, and the validity of feature is got over
It is high.In this patent, nine features have been selected, it can be ensured that the good description of image information.
In step (3), building significantization model generates the specific steps of initial notable figure are as follows:
The significance value of each super-pixel is obtained based on global area contrast and spatial relationship, and calculation is such as
Under:
Here pos (si,sj) indicate conspicuousness siAnd sjThe distance between.c(si) calculate pixel (xi,yi) and image
Space length between center (x ', y ') coordinate.vxAnd vyIt is the variable determined by the horizontal and vertical information of image.
In step (4), initial conspicuousness seed and final optimization pass conspicuousness are recursively refined using two cost functions
Figure, specific steps are as follows:
Conspicuousness s first by calculating each super-pixeliSmap is expressed as to obtain initial Saliency mapsk, k=0；So
Initial Saliency maps are divided into non-significant and marking area using the threshold value of Otsu afterwards.Non-significant and marking area can be regarded
For original image background seed (being expressed as BS) and foreground seeds (being expressed as FS).Due to the difference between super-pixel and background
Bigger, the significance value of super-pixel is higher.Bigger with the difference of prospect on the contrary, significance value is lower.Therefore, super-pixel is significant
Property value siIt can be recalculated based on BS and FS:
The process completes an iteration optimization, and forms a new Saliency maps Smapk, k=1.Next, weight
New BS and FS, follow-on Saliency maps Smap are newly obtained using the method for Otsuk+1It can be obtained by formula (4-6).
Determine whether iteration has reached termination condition by defining two cost functions.
Function f1(k) global fitness is measured, it means that the difference between a new generation and previous generation notable figure is smaller, mesh
Mark is just more acurrate.Function f2(k) local fitness is measured, it means that the variation between super-pixel super-pixel adjacent thereto is got over
Small, the significance value of each decision variable is bigger.By minimizing f1(k) and f2(k), optimal super-pixel grade can be generated
Notable figure.
The present invention is effective as follows:
Model proposed by the invention can realize state-of-the-art performance in soft image.
The present invention extracts foreground and background seed by recurrence and carries out conspicuousness calculating.In best features and conspicuousness seed
Guidance under, obtained Saliency maps can be generated by integrating multiple super-pixel grade Saliency maps on three scales.
The experimental results showed that the model proposed has been more than 15 progressive dies at first in three common data sets and nighttime image data set
Type, and there is optimum performance.
Detailed description of the invention
The basic flow chart of Fig. 1 the method for the present invention.
Fig. 2 is using the method for the present invention and existing image significance detection method respectively in MSRA data set, SOD data
The Saliency maps that the 16 conspicuousness models tested on collection, DUT-OMRON data set and NI data set proposed by the present invention obtain
Visual performance comparison diagram.
Specific embodiment
The embodiment of technical solution of the present invention is described in further detail with reference to the accompanying drawing.
1. as shown in Figure 1, dividing the image into super-pixel to retain and using the information of image；
2. as shown in Figure 1, extracting several visual signatures and therefrom selecting optimal feature；
3. as shown in Figure 1, building significantization model generates initial notable figure；
4. as shown in Figure 1, recursively refining initial conspicuousness seed and optimizing Saliency maps.
Step (1) is to retain and utilize the information of image, divides the image into super-pixel；
Step (2) extracts several visual signatures and therefrom selects optimal feature；
Step (3) constructs conspicuousness model and generates initial notable figure；
Step (4) recursively refines initial conspicuousness seed and optimizes Saliency maps.
In step (1), for object of reservation structural information and using the intermediate information of original image, changed using simple linear
Input picture is divided into super-pixel by generation cluster (SLIC) algorithm, is expressed as { si, i=1, N.The processing can lead to
It crosses and super-pixel is considered as processing unit to improve the efficiency of model.Since the accuracy of testing result is highly dependent on super-pixel
Quantity, the model proposed capture the super-pixel of three kinds of different scales, and wherein N is respectively set to 100,200 and 300.
In step (2), extract several lower-level vision features from input picture, including color characteristic, textural characteristics,
Direction character and Gradient Features, wherein color characteristic has nine kinds；The validity of lower-level vision feature is according to the difference of input picture
Contrast and change, the therefrom adaptively selected coldest days of the year end optimal characteristics of the comentropy based on lower-level vision feature.Lower-level vision is special
Sign extraction process is described as follows:
2-1. is first normalized the RGB color of input picture, to eliminate the influence of light and shade.Then will
Image is converted to LAB, HSV and YCbCr color space after normalization, to extract nine kinds of color characteristics.
2-2. indicates its textural characteristics using the two-dimensional entropy of image.The variation of textural characteristics is by the variation of entropy come really
Fixed, textural characteristics have the ability of very strong antinoise and geometry deformation.
2-3. is filtered by executing the Gabor of different directions θ ∈ { 0 °, 45 °, 90 °, 135 ° } on the gray level image of input
Device obtains direction character.It is smaller that the global property and rotational invariance of direction character influence it by low contrast.
2-4. calculates Gradient Features by average level gradient and vertical gradient.Part can be described by Gradient Features
The amplitude of grey scale change.
After the 1 dimension entropy by calculating each characteristic pattern, from extract 12 features L, A, B, H, S, V, Y, Cb, Cr,
E, O, G } in select nine optimal characteristics, nine optimal characteristics are expressed as { Fk, k=1,2 ..., 9 is its method are as follows:
Here pIIndicate the ratio I of the pixel where gray value.As characteristic statistics form, the polymerization of image grayscale distribution
The average information for including in characteristic can be reflected by image entropy.The Image entropy of characteristic pattern is bigger, and the validity of feature is got over
It is high.In this patent, nine features have been selected, it can be ensured that the good description of image information.
In step (3), building significantization model generates the specific steps of initial notable figure are as follows:
The significance value of each super-pixel is obtained based on global area contrast and spatial relationship, and calculation is such as
Under:
Here pos (si,sj) indicate conspicuousness siAnd sjThe distance between.c(si) calculate pixel (xi,yi) and image
Space length between center (x ', y ') coordinate.vxAnd vyIt is the variable determined by the horizontal and vertical information of image.
In step (4), initial conspicuousness seed and final optimization pass conspicuousness are recursively refined using two cost functions
Figure, specific steps are as follows:
Conspicuousness s first by calculating each super-pixeliSmap is expressed as to obtain initial Saliency mapsk, k=0；So
Initial Saliency maps are divided into non-significant and marking area using the threshold value of Otsu afterwards.Non-significant and marking area can be regarded
For original image background seed (being expressed as BS) and foreground seeds (being expressed as FS).Due to the difference between super-pixel and background
Bigger, the significance value of super-pixel is higher.Bigger with the difference of prospect on the contrary, significance value is lower.Therefore, super-pixel is significant
Property value siIt can be recalculated based on BS and FS:
The process completes an iteration optimization, and forms a new Saliency maps Smapk, k=1.Next,
The method of Otsu is reused to obtain new BS and FS, follow-on Saliency maps Smapk+1It can be obtained by formula (4-6)
?.Determine whether iteration has reached termination condition by defining two cost functions.
Function f1(k) global fitness is measured, it means that the difference between a new generation and previous generation notable figure is smaller, mesh
Mark is just more acurrate.Function f2(k) local fitness is measured, it means that the variation between super-pixel super-pixel adjacent thereto is got over
Small, the significance value of each decision variable is bigger.By minimizing f1(k) and f2(k), optimal super-pixel grade can be generated
Notable figure.
The method of the present invention and existing saliency detection model are respectively in MSRA data set, SOD data set, DUT-
The detection effect comparison of obtained notable figure is tested on OMRON data set and NI data set as shown in Fig. 2, wherein 1) MSRA number
Include 10000 natural images according to collection, there is simple background and high contrast；2) SOD data set includes multiple objects and answers
The image of miscellaneous background；3) DUT-OMRON data set includes relative complex and challenging image；4) our NI data set
The soft image shot comprising 200 evenings with vertical camera.The resolution ratio of every picture is 640 × 480, is additionally provided
The benchmark notable figure of hand labeled.It includes: IT model, SR model, FT model, NP model, CA that 15, which are made comparison conspicuousness model,
Model, IS model, LR model, PD model, MR model, SO model, BL model, GP model, SC model, SMD model, MIL model
And the model of the method for the present invention.All experiments make on Intel i5-5250CPU (1.6GHz) PC with 8GB RAM
It is executed with MATLAB.
We have used seven standards, including PR (precision-recall) curve, TPR-FPR to assess performance
(true positive rate and false positive rate) curve, AUC (area under the curve)
Point, MAE (mean absolute error) score, WF (weighted F-measure) score, OR (overlapping
Ratio) the average performance times (second) of score and each image.
It is compared by different threshold values by notable figure binaryzation and by itself and benchmark notable figure, so that it may obtain different
Precision P, recall rate R, kidney-Yang rate TPR and false positive rate FPR, PR curve and TPR-FPR song can be drawn by having obtained these ratios
Line.AUC score is the percentage of TPR-FPR area under the curve, and it is really aobvious that it can intuitively show that Saliency maps can be predicted
Write the fine or not degree of object.MAE score is considered as the mean absolute difference between Saliency maps obtained and true notable figure.Its
It is worth smaller, similitude is higher.F-measure score is defined as the weighted harmonic mean value between precision and recall rate, WF
Score is calculated by introducing weighting function to detection error.OR score be considered as binary saliency map and true notable figure it
Between the significant pixel of overlapping ratio.
Quantitative detection performance such as table 1 institute of the conspicuousness model proposed on four data sets with other 15 models
Show.Respectively with red, blue and green font highlight three best results in table 1.Indicate that the value is bigger to upward arrow
Performance is better, and down arrow indicates opposite meaning.The result shows that in most cases, the model proposed is in three public affairs
It ranks the first on image data set altogether or second, and real on soft image data set in relatively short time loss
Existing optimum performance.
Table 1. uses the quantitative result of five standard each conspicuousness models on four data sets
On MSRA and DUT-OMRON data set (Fig. 2), the model that is proposed TPRs-FPRs curve, PR curve and
Optimum performance is obtained on AUC score, and SO model reaches most preferably on MAE and WF score, MIL model reaches on OR score
Most preferably.This is because SO model improves robustness using boundary connectivity and global optimization, and to propose one kind more for MIL model
Case-based learning strategy improves accuracy.The two models are measured using background, and the significant object of complex background is effectively detected.
Although we are slightly below SO and MIL model at MAE, WF and the OR value of model, other scores of our score ratio are more competitive.
The average time-consuming of MIL model is more than every image 100 seconds, this is very inefficient.
On SOD data set (Fig. 2), the model proposed realizes optimality on TPR-FPR, PR, AUC, WF and OR
Energy.In terms of MAE scoring, the performance of our model be ranked second, and there was only little difference with the optimum of SO model
(0.002)。
On NI data set (Fig. 2), the model proposed is superior, is disappeared because it obtains the time in these standards
Consume relatively low optimum.
The qualitative comparison of the Saliency maps of 16 conspicuousness models is as shown in Fig. 2, the model proposed on four data sets
The true significant object in complicated and soft image can accurately be extracted.IT, NP, IS and SC model are easy by background
The influence of noise.Significant object can not be accurately positioned in SR and FT model.CA and PD model can not reflect the internal junction of significant object
Structure.The subjective performance of LR, MR, BL and GP model is influenced very big by low contrast background.SO, SMD and MIL model can not be
Obvious object is steadily detected under low contrast environment.The model proposed can be realized state-of-the-art in soft image
Performance.
It is to sum up told, the present invention constructs a kind of soft image conspicuousness detection side based on optimal feature selection
Method extracts foreground and background seed by recurrence and carries out conspicuousness calculating.It, can under the guidance of best features and conspicuousness seed
To generate obtained Saliency maps by integrating multiple super-pixel grade Saliency maps on three scales.Experimental result table
Bright, the model proposed has been more than 15 most advanced models in three common data sets and nighttime image data set, and is had
Optimum performance.
Finally, it should be noted that the foregoing is only a preferred embodiment of the present invention, it is not intended to restrict the invention,
While in accordance with previous embodiment, invention is explained in detail, for those skilled in the art, still can be with
It modifies to technical solution documented by previous embodiment or equivalent replacement of some of the technical features.It is all
Within the spirit and principles in the present invention, any modification, equivalent replacement, improvement and so on should be included in protection of the invention
Within the scope of.
The foregoing describe basic principles and main features of the invention and advantages of the present invention.Industry technical staff answers
The understanding, the present invention is not limited to the above embodiments, and the above embodiments and description only describe of the invention
Principle, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these change and change
Into all fall within the protetion scope of the claimed invention.
Claims (4)
1. a kind of soft image conspicuousness detection method based on optimal feature selection, it is characterised in that including walking as follows
It is rapid:
Step (1) is to retain and utilize the information of input picture, and input picture is divided into super-pixel；
Step (2) extracts several visual signatures and therefrom selects optimal feature；
Step (3) constructs conspicuousness model and generates initial notable figure；
Step (4) recursively refines initial conspicuousness seed and optimizes Saliency maps.
2. a kind of soft image conspicuousness detection method based on optimal feature selection according to claim 1,
It is characterized in that:
It is poly- using simple linear iteration for object of reservation structural information and using the intermediate information of input picture in step (1)
Input picture is divided into super-pixel by class algorithm, is expressed as { si }, i=1, N；Since the accuracy of testing result is high
Degree depends on the quantity of super-pixel, and wherein N is respectively set to 100,200 and 300；
Several lower-level vision features, including color characteristic, textural characteristics, direction are extracted in step (2) from input picture
Feature and Gradient Features, wherein color characteristic has nine kinds；The validity of lower-level vision feature is compared according to the difference of input picture
Degree and change, therefrom adaptively selected nine optimal characteristics of the comentropy based on lower-level vision feature, lower-level vision feature extraction
Process description is as follows:
2-1. is first normalized the RGB color of input picture, and image after normalization is then converted to LAB, HSV
With YCbCr color space, to extract nine kinds of color characteristics；
2-2. indicates its textural characteristics using the two-dimensional entropy of image, and the variation of textural characteristics is determined by the variation of entropy；
2-3. is come by executing the Gabor filter of different directions θ ∈ { 0 °, 45 °, 90 °, 135 ° } on the gray level image of input
Obtain direction character；
2-4. calculates Gradient Features by average level gradient and vertical gradient；
After the 1 dimension entropy by calculating each characteristic pattern, from 12 features { L, A, B, H, S, V, Y, Cb, Cr, E, O, G } extracted
Nine optimal characteristics of middle selection, nine optimal characteristics are expressed as { Fk, k=1,2 ..., 9 is its method are as follows:
Here pIIndicate the ratio I of the pixel where gray value；As characteristic statistics form, the polymerization property of image grayscale distribution
In include average information can be reflected by image entropy；The Image entropy of characteristic pattern is bigger, and the validity of feature is higher.
3. a kind of soft image conspicuousness detection method based on optimal feature selection according to claim 2,
It is characterized in that:
The specific implementation steps are as follows for step (3):
The significance value of each super-pixel is obtained based on global area contrast and spatial relationship, and calculation is as follows:
Here pos (si,sj) indicate conspicuousness siAnd sjThe distance between；c(si) calculate pixel (xi,yi) and picture centre
Space length between (x ', y ') coordinate；vxAnd vyIt is the variable determined by the horizontal and vertical information of image.
4. a kind of soft image conspicuousness detection method based on optimal feature selection according to claim 3,
It is characterized in that:
Initial conspicuousness seed and final optimization pass Saliency maps, tool are recursively refined in step (4) using two cost functions
Body step are as follows:
Conspicuousness s first by calculating each super-pixeliSmap is expressed as to obtain initial Saliency mapsk, k=0；Then make
Initial Saliency maps are divided into non-significant and marking area with the threshold value of Otsu；Non-significant and marking area can be considered as former
Beginning image background seed BS and foreground seeds FS, the significance value s of super-pixeliIt is recalculated based on BS and FS are as follows:
This recalculates process and completes an iteration optimization, and forms a new Saliency maps Smapk, k=1；It connects down
Come, reuses the method for Otsu to obtain new BS and FS, follow-on Saliency maps Smapk+1Pass through recurring formula (4-6)
It obtains；
Determine whether iteration has reached termination condition by defining two cost functions；
Function f1(k) global fitness is measured, it means that the difference between a new generation and previous generation notable figure is smaller, and target is just
It is more acurrate；Function f2(k) local fitness is measured, it means that the variation between super-pixel super-pixel adjacent thereto is smaller, often
The significance value of a decision variable is bigger；By minimizing f1(k) and f2(k) optimal super-pixel grade notable figure is generated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910086101.8A CN109886267A (en) | 2019-01-29 | 2019-01-29 | A kind of soft image conspicuousness detection method based on optimal feature selection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910086101.8A CN109886267A (en) | 2019-01-29 | 2019-01-29 | A kind of soft image conspicuousness detection method based on optimal feature selection |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109886267A true CN109886267A (en) | 2019-06-14 |
Family
ID=66927253
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910086101.8A Pending CN109886267A (en) | 2019-01-29 | 2019-01-29 | A kind of soft image conspicuousness detection method based on optimal feature selection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109886267A (en) |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110390327A (en) * | 2019-06-25 | 2019-10-29 | 北京百度网讯科技有限公司 | Foreground extracting method, device, computer equipment and storage medium |
| CN110991547A (en) * | 2019-12-12 | 2020-04-10 | 电子科技大学 | Image significance detection method based on multi-feature optimal fusion |
| CN111008585A (en) * | 2019-11-29 | 2020-04-14 | 西安电子科技大学 | Ship target detection method based on self-adaptive layered high-resolution SAR image |
| CN111160180A (en) * | 2019-12-16 | 2020-05-15 | 浙江工业大学 | Night green apple identification method of apple picking robot |
| CN111505038A (en) * | 2020-04-28 | 2020-08-07 | 中国地质大学（北京） | Implementation method for quantitatively analyzing sandstone cementation based on cathodoluminescence technology |
| CN111583279A (en) * | 2020-05-12 | 2020-08-25 | 重庆理工大学 | Super-pixel image segmentation method based on PCBA |
| TWI739401B (en) * | 2020-04-22 | 2021-09-11 | 國立中央大學 | Object classification method and object classification device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056590A (en) * | 2016-05-26 | 2016-10-26 | 重庆大学 | Manifold Ranking-based foreground- and background-characteristic combined saliency detection method |
| CN107169487A (en) * | 2017-04-19 | 2017-09-15 | 西安电子科技大学 | The conspicuousness object detection method positioned based on super-pixel segmentation and depth characteristic |
| US20170351941A1 (en) * | 2016-06-03 | 2017-12-07 | Miovision Technologies Incorporated | System and Method for Performing Saliency Detection Using Deep Active Contours |
| CN107992874A (en) * | 2017-12-20 | 2018-05-04 | 武汉大学 | Image well-marked target method for extracting region and system based on iteration rarefaction representation |
-
2019
- 2019-01-29 CN CN201910086101.8A patent/CN109886267A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106056590A (en) * | 2016-05-26 | 2016-10-26 | 重庆大学 | Manifold Ranking-based foreground- and background-characteristic combined saliency detection method |
| US20170351941A1 (en) * | 2016-06-03 | 2017-12-07 | Miovision Technologies Incorporated | System and Method for Performing Saliency Detection Using Deep Active Contours |
| CN107169487A (en) * | 2017-04-19 | 2017-09-15 | 西安电子科技大学 | The conspicuousness object detection method positioned based on super-pixel segmentation and depth characteristic |
| CN107992874A (en) * | 2017-12-20 | 2018-05-04 | 武汉大学 | Image well-marked target method for extracting region and system based on iteration rarefaction representation |
Non-Patent Citations (3)
| Title |
|---|
| MU NAN等: "A Multiscale Superpixel-level Salient Object Detection Model Using Local-Global Contrast Cue", 《JOURNAL OF SHANGHAI JIAOTONG UNIVERSITY (SCIENCE)》 * |
| NAN MU等: "Optimal Feature Selection for Saliency Seed Propagation in Low Contrast Images", 《ADVANCES IN MULTIMEDIA INFORMATION PROCESSING – PCM 2018》 * |
| 张艳邦等: "基于颜色和纹理特征的显著性检测算法", 《计算机应用研究》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110390327A (en) * | 2019-06-25 | 2019-10-29 | 北京百度网讯科技有限公司 | Foreground extracting method, device, computer equipment and storage medium |
| CN110390327B (en) * | 2019-06-25 | 2022-06-28 | 北京百度网讯科技有限公司 | Foreground extraction method and device, computer equipment and storage medium |
| CN111008585A (en) * | 2019-11-29 | 2020-04-14 | 西安电子科技大学 | Ship target detection method based on self-adaptive layered high-resolution SAR image |
| CN111008585B (en) * | 2019-11-29 | 2023-04-07 | 西安电子科技大学 | Ship target detection method based on self-adaptive layered high-resolution SAR image |
| CN110991547A (en) * | 2019-12-12 | 2020-04-10 | 电子科技大学 | Image significance detection method based on multi-feature optimal fusion |
| CN111160180A (en) * | 2019-12-16 | 2020-05-15 | 浙江工业大学 | Night green apple identification method of apple picking robot |
| TWI739401B (en) * | 2020-04-22 | 2021-09-11 | 國立中央大學 | Object classification method and object classification device |
| CN111505038A (en) * | 2020-04-28 | 2020-08-07 | 中国地质大学（北京） | Implementation method for quantitatively analyzing sandstone cementation based on cathodoluminescence technology |
| CN111583279A (en) * | 2020-05-12 | 2020-08-25 | 重庆理工大学 | Super-pixel image segmentation method based on PCBA |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109886267A (en) | A kind of soft image conspicuousness detection method based on optimal feature selection | |
| CN111062973B (en) | Vehicle tracking method based on target feature sensitivity and deep learning | |
| CN107016405B (en) | A kind of pest image classification method based on classification prediction convolutional neural networks | |
| CN102214298B (en) | Method for detecting and identifying airport target by using remote sensing image based on selective visual attention mechanism | |
| CN104077605B (en) | A kind of pedestrian's search recognition methods based on color topological structure | |
| Hsu et al. | An interactive flower image recognition system | |
| CN109800629A (en) | A kind of Remote Sensing Target detection method based on convolutional neural networks | |
| CN106649487A (en) | Image retrieval method based on interest target | |
| CN113963222B (en) | High-resolution remote sensing image change detection method based on multi-strategy combination | |
| CN105528575B (en) | Sky detection method based on Context Reasoning | |
| CN104881671B (en) | A kind of high score remote sensing image Local Feature Extraction based on 2D Gabor | |
| CN109325484A (en) | Flowers image classification method based on background priori conspicuousness | |
| CN102496023A (en) | Region of interest extraction method of pixel level | |
| CN107633226A (en) | A kind of human action Tracking Recognition method and system | |
| CN108629783A (en) | Image partition method, system and medium based on the search of characteristics of image density peaks | |
| CN105493141A (en) | Unstructured road boundary detection | |
| CN108664838A (en) | Based on the monitoring scene pedestrian detection method end to end for improving RPN depth networks | |
| CN111860587B (en) | Detection method for small targets of pictures | |
| CN108596195B (en) | Scene recognition method based on sparse coding feature extraction | |
| CN107808376A (en) | A kind of detection method of raising one's hand based on deep learning | |
| CN114693616B (en) | Rice disease detection method, device and medium based on improved target detection model and convolutional neural network | |
| CN111199245A (en) | Rape pest identification method | |
| CN106446925A (en) | Dolphin identity recognition method based on image processing | |
| CN106650798A (en) | Indoor scene recognition method combining deep learning and sparse representation | |
| CN114495170A (en) | Pedestrian re-identification method and system based on local self-attention inhibition |
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: 20190614 |