CN107886507B - A kind of salient region detecting method based on image background and spatial position - Google Patents

A kind of salient region detecting method based on image background and spatial position Download PDF

Info

Publication number
CN107886507B
CN107886507B CN201711122796.8A CN201711122796A CN107886507B CN 107886507 B CN107886507 B CN 107886507B CN 201711122796 A CN201711122796 A CN 201711122796A CN 107886507 B CN107886507 B CN 107886507B
Authority
CN
China
Prior art keywords
super
pixel block
cluster
pixel
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201711122796.8A
Other languages
Chinese (zh)
Other versions
CN107886507A (en
Inventor
王慧
刘钢
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Changchun University of Technology
Original Assignee
Changchun University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Changchun University of Technology filed Critical Changchun University of Technology
Priority to CN201711122796.8A priority Critical patent/CN107886507B/en
Publication of CN107886507A publication Critical patent/CN107886507A/en
Application granted granted Critical
Publication of CN107886507B publication Critical patent/CN107886507B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • 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
    • G06T2207/20021—Dividing image into blocks, subimages or windows

Abstract

The present invention provides a kind of salient region detecting method based on image background and spatial position, and method mainly includes the following steps：The super-pixel that M kind scales are carried out to target image is layered segmentation, obtains M layers of target subgraph；Super-pixel block characteristic vector pickup extracts the color characteristic and textural characteristics of super-pixel block；Background super-pixel block clusters；Conspicuousness calculating is carried out using spatial position and with background super-pixel difference to each super-pixel block；Multiple dimensioned super-pixel block conspicuousness fusion.Advantage is：The method of the invention is capable of the marking area of accurate judgement image, can effectively improve the accuracy rate and computational efficiency of detection, for be applied to internet, cloud computing large nuber of images or video data screening and analysis technical support is provided.

Description

A kind of salient region detecting method based on image background and spatial position
Technical field
The invention belongs to saliency region detection technical fields, and in particular to one kind being based on image background and space bit The salient region detecting method set.
Background technology
For the angle for pushing high-grade intelligent robot to research and develop, marking area detection can make intelligent robot from same In the multitude of video data that time receives, filters out and handled with the maximally related part of current task.This can effective mould The directive property and centrality of quasi- human visual perception lay the foundation to complete intelligent task.From promotion visual field intelligent use Angle for, by salient region detecting method be applied to internet, the large nuber of images of cloud computing or video data screening and In analysis, the accuracy rate and computational efficiency of detection can be effectively improved；It is applied to reconnaissance plane, field of video monitoring, can be that target is known Not, follow-up of hot issues scheduling algorithm offer key area early period label, improves the computational efficiency of related algorithm；It is applied to image or video Transmission field can targetedly compress the key area on image or video, improve the effect of image or transmission of video Rate.In addition, salient region detecting method also can be widely used to other fields such as path navigation, unmanned plane.
In recent years, numerous scholars propose many for detecting salient region or mesh calibration method in the picture.To carry Computationally efficient simultaneously ignores the unnecessary details of some in image, these methods extract the perception homogeneity member of image first mostly Element then calculates their local contrast, the overall situation such as super pixel, region (also having the method for directly using pixel certainly) Comparative or sparse noise is finally integrated with obtaining significance value of each perception with prime element to divide entire notable mesh Mark.From the point of view of research tendency in recent years, relative to local contrast, global clue can be divided due to it on similar picture areas It is more concerned equipped with comparative saliency value.
The patent of entitled " a kind of saliency method for detecting area and system ", 104424642 A of Publication No. CN A kind of saliency method for detecting area and system are disclosed, by the static significant characteristics, the part that obtain Pixel-level respectively The static significant characteristics of region class, the dynamic significant characteristics of regional area grade, the static significant characteristics of global level and complete The dynamic significant characteristics of office's grade, are modulated the saliency feature using the correlation between video frame, based on tune The saliency region of video frame is arranged using 3D-MRF for saliency feature after system, then utilizes Graph-cuts Optimal saliency region is selected, saliency region is split.This method application marking area detects mutual Benefit property priori improves the performance of algorithm, but when the borderline region of image well cannot describe background, such as frame region spy When sign differs greatly, entire frame is put together and calculates background characteristics, this method is inaccurate to the calculating of background characteristics.
The patent of entitled " a kind of detection method of salient region " discloses a kind of detection method of salient region, The basic element for participating in otherness comparing calculation is defined as region by it, is allowed to final testing result in same magnitude, from And improve the efficiency of salient region detection.But the invention only apply color space conversion and figure segmentation etc. part it is right Than degree, when image object unobvious, effect is bad.
The patent of entitled " a kind of saliency method for detecting area of deep learning " discloses a kind of deep learning Saliency method for detecting area obtain image by the way that the result of heterogeneous networks layer under deep learning to be combined and exist Feature under different scale, to obtain better detection performance；Simultaneously super-pixel threshold learning is carried out using image segmentation.But It is image category (complex background or simple background, including simple target or multiple of the method by its training set of invention proposition Target) and quantity influence, this method, which is susceptible to, is excessively applicable in risk, may be showed not when image category changes It is good.
It can be seen that above-mentioned all kinds of saliency method for detecting area, all have certain use limitation, to lead Cause the accuracy rate of detection not high, the algorithm of detection is excessively complicated.
Invention content
In view of the defects existing in the prior art, the present invention provides a kind of marking area based on image background and spatial position Detection method can effectively solve the above problems.
The technical solution adopted by the present invention is as follows：
The present invention provides a kind of salient region detecting method based on image background and spatial position, includes the following steps：
Step 1, the super-pixel for M kind scales being carried out to target image is layered segmentation, wherein M is total number of plies of scale, is obtained M layers of target subgraph；Every layer of target subgraph is made of multiple super-pixel block；
Step 2, for every layer of target subgraph, following steps 2.1- steps 2.3 are performed both by：
Step 2.1, the feature vector for extracting each super-pixel block in target subgraph, obtains super-pixel block feature vector；
Step 2.2, by the frame region of target subgraph as image background, the super-pixel block for belonging to image background is known as Background super-pixel block；
Background super-pixel block is clustered, obtains n cluster, respectively：1st cluster, the 2nd clusters ... n-th Cluster；The cluster centre feature vector of 1st cluster is B1, the cluster centre feature vector of the 2nd cluster is B2, and so on, The cluster centre feature vector of n-th of cluster is Bn, therefore, cluster centre feature vector B={ B1,B2..., Bn}；
Step 2.3, for each super-pixel block of target subgraph, it is expressed as super-pixel block p, it is super to be all made of following formula calculating The significance value s of block of pixels p：
Wherein：
Wherein：
D(p,Bi) indicate super-pixel block p and ith cluster cluster centre feature vector BiThe distance between, i=1, 2 ..., n }；σ represents scale factor；
W is weights, for weigh super-pixel block p between this layer of target subgraph central point at a distance from, (x, y) indicates super picture The center point coordinate of plain block p, (x', y') indicate the center point coordinate of this layer of target subgraph；
Thus the significance value of each super-pixel block of every layer of target subgraph is calculated；
Step 3, multiple dimensioned super-pixel block conspicuousness fusion, obtains final Saliency maps, and detected on Saliency maps To salient region, specifically include：
Step 3.1, the significance value of arbitrary pixel j on Saliency maps after merging is calculated:
The significance value s of pixel jjIt is its being averaged positioned at the significance value of corresponding super-pixel block under all scales Value, i.e.,：
Wherein：slIt is the significance value for the super-pixel block that pixel j is located at l layers of target subgraph；
Step 3.2, the significance value of all pixels point j forms image saliency map, is more than setting threshold on Saliency maps The region of value is the salient region eventually detected.
Preferably, in step 1, the super-pixel for carrying out M kind scales to target image using SLIC algorithms is layered segmentation.
Preferably, in step 2.1, the feature vector of each super-pixel block in target subgraph is extracted, specially：Extraction is every The color characteristic and textural characteristics of a super-pixel block, the feature vector of each super-pixel block include：3 components of RGB average values, 3 points of 256 components of RGB histograms, 3 components of HSV average values, 256 components of HSV histograms, Lab average values 48 components of amount, 256 components of Lab histograms and the response of LM filters.
Preferably, in step 2.2, background super-pixel block is clustered, specifically, being calculated using K-Means clusters are improved Method clusters background super-pixel block.
Preferably, background super-pixel block is clustered using improvement K-Means clustering algorithms, is specifically included：
Step 2.2.1 sets the initial clustering number of improved K-means clustering algorithms as z, i.e., last cluster obtains z Cluster numbers；
Step 2.2.2 carries out initial clustering using K-means clustering algorithms, obtains several initial clusterings；Initially gathering When class, the distance of any two super-pixel block is calculated using following methods：
For any two super-pixel block in target subgraph, it is denoted as super-pixel block u and super-pixel block v respectively；
If the RGB average values for extracting super-pixel block u in target subgraph are f1 u, RGB histograms areHSV average values areHSV histograms areLab average values areLab histograms areLM filters respond
If the RGB average values for extracting super-pixel block v in target subgraph are f1 v, RGB histograms areHSV average values areHSV histograms areLab average values areLab histograms areLM filters respond
The distance between super-pixel block u and super-pixel block v D (u, v) is：
Wherein：N (●) indicates normalization；
Indicate the distance of a-th of feature between super-pixel block u and super-pixel block v；
Wherein, a=1,3,5,7, it respectively represents RGB and is averaged value tag, HSV is averaged value tag, and Lab is averaged value tag and LM Filter response characteristic；M is the dimension sum of each feature, and e is the number of dimensions parameter of each feature, is averaged value tag for RGB, Its dimension is 3；It is averaged value tag for HSV, dimension 3；It is averaged value tag for Lab, dimension 3；LM is filtered Device response characteristic, dimension 48；It is e-th of component of a-th of feature of super-pixel block u；It is super-pixel block v A-th of feature e-th of component；
Indicate the distance of c-th of feature between super-pixel block u and super-pixel block v；Wherein, c =2,4,6, respectively represent RGB histograms, HSV histograms and Lab histograms；B is histogram number；D is histogram area Between number parameter；It is d-th of histogram value of c-th of feature of super-pixel block u；It is c-th of feature of super-pixel block v D-th of histogram value；
Then the cluster centre feature vector of each initial clustering is calculated；By the feature of all super-pixel in a cluster It does respectively and is averagely worth to cluster centre；
Step 2.2.3, selectes Euclidean distance as the similarity measurement between initial clustering, to calculate cluster centre it Between difference value；
Step 2.2.4, judges whether the difference of any two cluster centre is less than threshold θ；If cluster centre collection is combined into A, ThenD (g, h) represents the Euclidean distance between cluster centre g and cluster centre h；
Step 2.2.5, if the result of step 2.2.4 is "Yes", the number N clustered subtracts 1, return to step 2.2.2 weights New cluster；
Step 2.2.6 enters step 2.2.7 if the result of step 2.2.4 is "No"；
Step 2.2.7 records the number of cluster and the feature vector of cluster centre；
Step 2.2.8, flow terminate.
A kind of salient region detecting method based on image background and spatial position provided by the invention has the following advantages：
The method of the invention is capable of the marking area of accurate judgement image, and expression effect is good, can effectively improve detection Accuracy rate and computational efficiency, for be applied to internet, cloud computing large nuber of images or video data screening and analysis provide Technical support.
Description of the drawings
Fig. 1 is that the overall flow of the salient region detecting method provided by the invention based on image background and spatial position is shown It is intended to；
Fig. 2 is that the super-pixel of the salient region detecting method provided by the invention based on image background and spatial position is layered Segmentation result schematic diagram；
Fig. 3 is a kind of method flow diagram improving K-means clustering algorithms provided by the invention；
Fig. 4 is marking area testing result contrast schematic diagram.
Specific implementation mode
In order to make the technical problems, technical solutions and beneficial effects solved by the present invention be more clearly understood, below in conjunction with Accompanying drawings and embodiments, the present invention will be described in further detail.It should be appreciated that specific embodiment described herein only to It explains the present invention, is not intended to limit the present invention.
The present invention provides a kind of salient region detecting method based on image background and spatial position, and method includes mainly such as Lower step：The super-pixel that M kind scales are carried out to target image is layered segmentation, obtains M layers of target subgraph；Super-pixel block feature to Amount extraction, extracts the color characteristic and textural characteristics of super-pixel block；Background super-pixel block clusters；Utilize spatial position and and background Super-pixel difference carries out conspicuousness calculating to each super-pixel block；Multiple dimensioned super-pixel block conspicuousness fusion.Side of the present invention Method is capable of the marking area of accurate judgement image, and expression effect is good, can effectively improve the accuracy rate and computational efficiency of detection, be It is applied to internet, the screening of the large nuber of images of cloud computing or video data and analysis and technical support is provided.
Salient region detecting method based on image background and spatial position carries out the super picture of image background on every tomographic image Element cluster, calculates super-pixel feature vector, and according to the difference of super-pixel spatial position and itself and background super-pixel, it is super to calculate this Pixel significance value.Finally super-pixel saliency value on each tomographic image is merged, obtains final notable figure.With reference to figure 1, including Following steps：
Step 1, the super-pixel for M kind scales being carried out to target image is layered segmentation, wherein M is total number of plies of scale, is obtained M layers of target subgraph；Every layer of target subgraph is made of multiple super-pixel block；
In this step, specifically use SLIC (Simple Linear Iterative Cluster) algorithm to target image Carry out the super-pixel layering segmentation of M kind scales.Image superpixel layering segmentation can simulate the difference of human eye difference cellula visualis Vision granularity carries out conspicuousness judgement from different scale to obtain more good effect to pixel on image, is obtained to final To an objective notable figure of justice.Consider human eye feature and algorithm performance, 3 layers of super-pixel layering segmentation side can be taken Method.
Step 2, for every layer of target subgraph, following steps 2.1- steps 2.3 are performed both by：
Step 2.1, the feature vector for extracting each super-pixel block in target subgraph, obtains super-pixel block feature vector； Specifically, extracting the color characteristic and textural characteristics of each super-pixel block, the feature vector of each super-pixel block includes：RGB is flat 3 components of mean value, 256 components of RGB histograms, 3 components of HSV average values, HSV histograms 256 components, 48 components of 3 components of Lab average values, 256 components of Lab histograms and the response of LM filters.
Step 2.2, by the frame region of target subgraph as image background, the super-pixel block for belonging to image background is known as Background super-pixel block；
Background super-pixel block is clustered, obtains n cluster, respectively：1st cluster, the 2nd clusters ... n-th Cluster；The cluster centre feature vector of 1st cluster is B1, the cluster centre feature vector of the 2nd cluster is B2, and so on, The cluster centre feature vector of n-th of cluster is Bn, therefore, cluster centre feature vector B={ B1,B2..., Bn}；
Under normal circumstances, background super-pixel block is clustered to 1-3 cluster set, can be prevented because of framing mask super-pixel Background characteristics vector caused by difference is larger calculates mistake, to give one more accurate evaluation method of background node.
In this step, background super-pixel block is clustered, specifically, using K-Means clustering algorithms are improved to background Super-pixel block is clustered, and with reference to figure 3, is included the following steps：
Step 2.2.1 sets the initial clustering number of improved K-means clustering algorithms as z, i.e., last cluster obtains z Cluster numbers；General z values are 3；
Step 2.2.2 carries out initial clustering using K-means clustering algorithms, obtains several initial clusterings；Initially gathering When class, the distance of any two super-pixel block is calculated using following methods：
For any two super-pixel block in target subgraph, it is denoted as super-pixel block u and super-pixel block v respectively；
If the RGB average values for extracting super-pixel block u in target subgraph are f1 u, RGB histograms areHSV average values areHSV histograms areLab average values areLab histograms areLM filters respond
If the RGB average values for extracting super-pixel block v in target subgraph are f1 v, RGB histograms areHSV average values areHSV histograms areLab average values areLab histograms areLM filters respond
The distance between super-pixel block u and super-pixel block v D (u, v) is：
Wherein：N (●) indicates normalization；
Indicate the distance of a-th of feature between super-pixel block u and super-pixel block v；
Wherein, a=1,3,5,7, it respectively represents RGB and is averaged value tag, HSV is averaged value tag, and Lab is averaged value tag and LM Filter response characteristic；M is the dimension sum of each feature, and e is the number of dimensions parameter of each feature, is averaged value tag for RGB, Its dimension is 3；It is averaged value tag for HSV, dimension 3；It is averaged value tag for Lab, dimension 3；LM is filtered Device response characteristic, dimension 48；It is e-th of component of a-th of feature of super-pixel block u；It is super-pixel block v A-th of feature e-th of component；
Indicate the distance of c-th of feature between super-pixel block u and super-pixel block v；Wherein, c =2,4,6, respectively represent RGB histograms, HSV histograms and Lab histograms；B is histogram number；D is histogram area Between number parameter；It is d-th of histogram value of c-th of feature of super-pixel block u；It is c-th of feature of super-pixel block v D-th of histogram value；
Then the cluster centre feature vector of each initial clustering is calculated；By the feature of all super-pixel in a cluster It does respectively and is averagely worth to cluster centre；
Step 2.2.3, selectes Euclidean distance as the similarity measurement between initial clustering, to calculate cluster centre it Between difference value；
Step 2.2.4, judges whether the difference of any two cluster centre is less than threshold θ；If cluster centre collection is combined into A, ThenD (g, h) represents the Euclidean distance between cluster centre g and cluster centre h；
Step 2.2.5, if the result of step 2.2.4 is "Yes", the number N clustered subtracts 1, return to step 2.2.2 weights New cluster；
Step 2.2.6 enters step 2.2.7 if the result of step 2.2.4 is "No"；
Step 2.2.7 records the number of cluster and the feature vector of cluster centre；
Step 2.2.8, flow terminate.
In this step, background priori is based on physics of photography, and four frame regions of image are treated as image background.Mesh Preceding most of algorithms using background priori extract background area feature vector using the entire frame of image as background, this Mode cannot efficiently use framing mask background difference.The survey found that the region of many framing masks can be divided into 1 to 3 A part, and generally below 3 parts.Therefore, it is good description image background regions, super-pixel segmentation is being carried out to image On the basis of, for framing mask background super-pixel block set, the present invention is using k-means clustering algorithms are improved, by image four Background super-pixel block cluster is 1 to 3 set on a frame, as image background regions.Utilize the institute on four frames of image The super-pixel block that has powerful connections forms background super-pixel block set, extracts color characteristic and the textural characteristics of all super-pixel block to describe Super-pixel block information.
Step 2.3, for each super-pixel block of target subgraph, it is expressed as super-pixel block p, it is super to be all made of following formula calculating The significance value s of block of pixels p：
Wherein：
Wherein：
D(p,Bi) indicate super-pixel block p and ith cluster cluster centre feature vector BiThe distance between, i=1, 2 ..., n }；σ represents scale factor, and usual value is 0.5；
W is weights, for weigh super-pixel block p between this layer of target subgraph central point at a distance from, (x, y) indicates super picture The center point coordinate of plain block p, (x', y') indicate the center point coordinate of this layer of target subgraph；
Thus the significance value of each super-pixel block of every layer of target subgraph is calculated；
This step carry out super-pixel block significance value calculating when, using spatial position and with background super-pixel difference pair Super-pixel block carries out conspicuousness calculating, specially：Background super-pixel block cluster is carried out on every layer of target subgraph, according to super picture The difference of plain block space position and itself and background super-pixel block calculates the super-pixel block significance value.Super-pixel block p's is notable Property value is the weighted average of itself and had powerful connections super-pixel block cluster centre difference, weight and itself and the tomographic image central point Distance dependent, apart from smaller, weight is bigger.
Step 3, multiple dimensioned super-pixel block conspicuousness fusion, obtains final Saliency maps, and detected on Saliency maps To salient region, specifically include：
Step 3.1, the significance value of arbitrary pixel j on Saliency maps after merging is calculated:
The significance value s of pixel jjIt is its being averaged positioned at the significance value of corresponding super-pixel block under all scales Value, i.e.,：
Wherein：slIt is the significance value for the super-pixel block that pixel j is located at l layers of target subgraph；
Step 3.2, the significance value of all pixels point j forms image saliency map, is more than setting threshold on Saliency maps The region of value is the salient region eventually detected.
Using proposed by the present invention based on image background and the salient region detecting method BSP of spatial position, classic algorithm GR, classic algorithm SF carry out marking area detection to the original graph in Fig. 4 respectively, and testing result by result as shown in figure 4, illustrated Figure Fig. 4 is it is found that the marking area detection result of BSP algorithm of the present invention is good, hence it is evident that is better than classic algorithm GR and classic algorithm SF. In addition, area AUC under the mean absolute error MAE and ROC curve of three kinds of detection algorithms of calculating, result of calculation is as shown in the table, From following table as can be seen that the MAE values of BSP are less than GR and SF；The AUC value of BSP is higher than GR and SF, it is indicated above that BSP methods is comprehensive It is good to close performance.
Table：Three kinds of detection algorithms compare
A kind of salient region detecting method based on image background and spatial position provided by the invention has following excellent Point：
By the method for the invention, it is capable of the marking area of accurate judgement image, the accuracy rate and meter of detection can be effectively improved Calculate efficiency, for be applied to internet, cloud computing large nuber of images or video data screening and analysis technical support is provided, have Good application prospect.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For member, various improvements and modifications may be made without departing from the principle of the present invention, these improvements and modifications are also answered Depending on protection scope of the present invention.

Claims (5)

1. a kind of salient region detecting method based on image background and spatial position, which is characterized in that include the following steps：
Step 1, the super-pixel for M kind scales being carried out to target image is layered segmentation, wherein M is total number of plies of scale, obtains M layers Target subgraph；Every layer of target subgraph is made of multiple super-pixel block；
Step 2, for every layer of target subgraph, following steps 2.1- steps 2.3 are performed both by：
Step 2.1, the feature vector for extracting each super-pixel block in target subgraph, obtains super-pixel block feature vector；
Step 2.2, by the frame region of target subgraph as image background, the super-pixel block for belonging to image background is known as background Super-pixel block；
Background super-pixel block is clustered, obtains n cluster, respectively：1st cluster, n-th of cluster of the 2nd cluster ...； The cluster centre feature vector of 1st cluster is B1, the cluster centre feature vector of the 2nd cluster is B2, and so on, n-th The cluster centre feature vector of cluster is Bn, therefore, cluster centre feature vector B={ B1,B2..., Bn}；
Step 2.3, for each super-pixel block of target subgraph, it is expressed as super-pixel block p, following formula is all made of and calculates super-pixel The significance value s of block p：
Wherein：
Wherein：
D(p,Bi) indicate super-pixel block p and ith cluster cluster centre feature vector BiThe distance between, i=1,2 ..., n} ；σ represents scale factor；
W is weights, for weigh super-pixel block p between this layer of target subgraph central point at a distance from, (x, y) indicate super-pixel block The center point coordinate of p, (x', y') indicate the center point coordinate of this layer of target subgraph；
Thus the significance value of each super-pixel block of every layer of target subgraph is calculated；
Step 3, multiple dimensioned super-pixel block conspicuousness fusion, obtains final Saliency maps, and detects on Saliency maps aobvious Work property region, specifically includes：
Step 3.1, the significance value of arbitrary pixel j on Saliency maps after merging is calculated:
The significance value s of pixel jjIt is that it is located at the average value of the significance value of corresponding super-pixel block under all scales, I.e.：
Wherein：slIt is the significance value for the super-pixel block that pixel j is located at l layers of target subgraph；
Step 3.2, the significance value of all pixels point j forms image saliency map, is more than given threshold on Saliency maps Region is the salient region eventually detected.
2. a kind of salient region detecting method based on image background and spatial position according to claim 1, feature It is, in step 1, the super-pixel for carrying out M kind scales to target image using SLIC algorithms is layered segmentation.
3. a kind of salient region detecting method based on image background and spatial position according to claim 1, feature It is, in step 2.1, extracts the feature vector of each super-pixel block in target subgraph, specially：Extract each super-pixel block Color characteristic and textural characteristics, the feature vector of each super-pixel block includes：3 components, the RGB histograms of RGB average values 256 components, 3 components of HSV average values, 256 components of HSV histograms, 3 components, Lab of Lab average values it is straight 48 components of 256 components and LM the filters response of square figure.
4. a kind of salient region detecting method based on image background and spatial position according to claim 3, feature It is, in step 2.2, background super-pixel block is clustered, specifically, super to background using K-Means clustering algorithms are improved Block of pixels is clustered.
5. a kind of salient region detecting method based on image background and spatial position according to claim 4, feature It is, background super-pixel block is clustered using K-Means clustering algorithms are improved, is specifically included：
Step 2.2.1 sets the initial clustering number of improved K-means clustering algorithms as z, i.e., last cluster obtains z cluster Number；
Step 2.2.2 carries out initial clustering using K-means clustering algorithms, obtains several initial clusterings；In initial clustering When, the distance of any two super-pixel block is calculated using following methods：
For any two super-pixel block in target subgraph, it is denoted as super-pixel block u and super-pixel block v respectively；
If the RGB average values for extracting super-pixel block u in target subgraph are f1 u, RGB histograms areHSV average values are HSV histograms areLab average values areLab histograms areLM filters respond
If the RGB average values for extracting super-pixel block v in target subgraph are f1 v, RGB histograms areHSV average values are HSV histograms areLab average values areLab histograms areLM filters respond
The distance between super-pixel block u and super-pixel block v D (u, v) is：
Wherein：N (●) indicates normalization；
Indicate the distance of a-th of feature between super-pixel block u and super-pixel block v；
Wherein, a=1,3,5,7, it respectively represents RGB and is averaged value tag, HSV is averaged value tag, and Lab is averaged value tag and LM filtering Device response characteristic；M is the dimension sum of each feature, and e is the number of dimensions parameter of each feature, be averaged value tag, ties up for RGB Degree is 3；It is averaged value tag for HSV, dimension 3；It is averaged value tag for Lab, dimension 3；LM filters are rung Answer feature, dimension 48；It is e-th of component of a-th of feature of super-pixel block u；It is the of super-pixel block v E-th of component of a feature；
Indicate the distance of c-th of feature between super-pixel block u and super-pixel block v；Wherein, c=2, 4,6, respectively represent RGB histograms, HSV histograms and Lab histograms；B is histogram number；D is histogram number Measure parameter；It is the d histogram value of c-th of feature of super-pixel block u；It is c-th of feature of super-pixel block v D-th of histogram value；
Then the cluster centre feature vector of each initial clustering is calculated；By the feature difference of all super-pixel in a cluster It does and is averagely worth to cluster centre；
Step 2.2.3, selectes Euclidean distance as the similarity measurement between initial clustering, to calculate between cluster centre Difference value；
Step 2.2.4, judges whether the difference of any two cluster centre is less than threshold θ；If cluster centre collection is combined into A, thenD (g, h) represents the Euclidean distance between cluster centre g and cluster centre h；
Step 2.2.5, if the result of step 2.2.4 is "Yes", the number N clustered subtracts 1, and return to step 2.2.2 gathers again Class；
Step 2.2.6 enters step 2.2.7 if the result of step 2.2.4 is "No"；
Step 2.2.7 records the number of cluster and the feature vector of cluster centre；
Step 2.2.8, flow terminate.
CN201711122796.8A 2017-11-14 2017-11-14 A kind of salient region detecting method based on image background and spatial position Active CN107886507B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201711122796.8A CN107886507B (en) 2017-11-14 2017-11-14 A kind of salient region detecting method based on image background and spatial position

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201711122796.8A CN107886507B (en) 2017-11-14 2017-11-14 A kind of salient region detecting method based on image background and spatial position

Publications (2)

Publication Number Publication Date
CN107886507A CN107886507A (en) 2018-04-06
CN107886507B true CN107886507B (en) 2018-08-21

Family

ID=61776610

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201711122796.8A Active CN107886507B (en) 2017-11-14 2017-11-14 A kind of salient region detecting method based on image background and spatial position

Country Status (1)

Country Link
CN (1) CN107886507B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108921820B (en) * 2018-05-30 2021-10-29 咸阳师范学院 Saliency target detection method based on color features and clustering algorithm
CN110866896B (en) * 2019-10-29 2022-06-24 中国地质大学（武汉） Image saliency target detection method based on k-means and level set super-pixel segmentation
CN112017158B (en) * 2020-07-28 2023-02-14 中国科学院西安光学精密机械研究所 Spectral characteristic-based adaptive target segmentation method in remote sensing scene
CN112085020B (en) * 2020-09-08 2023-08-01 北京印刷学院 Visual saliency target detection method and device
CN112418147A (en) * 2020-12-02 2021-02-26 中国人民解放军军事科学院国防科技创新研究院 Track \30758identificationmethod and device based on aerial images
CN113378873A (en) * 2021-01-13 2021-09-10 杭州小创科技有限公司 Algorithm for determining attribution or classification of target object

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1770161A (en) * 2004-09-29 2006-05-10 英特尔公司 K-means clustering using t-test computation
CN105913456A (en) * 2016-04-12 2016-08-31 西安电子科技大学 Video significance detecting method based on area segmentation
CN106203430A (en) * 2016-07-07 2016-12-07 北京航空航天大学 A kind of significance object detecting method based on foreground focused degree and background priori

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7139433B2 (en) * 2003-03-13 2006-11-21 Sharp Laboratories Of America, Inc. Compound image compression method and apparatus

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1770161A (en) * 2004-09-29 2006-05-10 英特尔公司 K-means clustering using t-test computation
CN105913456A (en) * 2016-04-12 2016-08-31 西安电子科技大学 Video significance detecting method based on area segmentation
CN106203430A (en) * 2016-07-07 2016-12-07 北京航空航天大学 A kind of significance object detecting method based on foreground focused degree and background priori

Also Published As

Publication number Publication date
CN107886507A (en) 2018-04-06

Similar Documents

Publication Publication Date Title
CN107886507B (en) A kind of salient region detecting method based on image background and spatial position
CN104408429B (en) A kind of video represents frame extracting method and device
CN106570486B (en) Filtered target tracking is closed based on the nuclear phase of Fusion Features and Bayes's classification
CN106203430B (en) A kind of conspicuousness object detecting method based on foreground focused degree and background priori
CN110532970B (en) Age and gender attribute analysis method, system, equipment and medium for 2D images of human faces
CN104517095B (en) A kind of number of people dividing method based on depth image
US20160358035A1 (en) Saliency information acquisition device and saliency information acquisition method
CN107392968B (en) The image significance detection method of Fusion of Color comparison diagram and Color-spatial distribution figure
CN103020992B (en) A kind of video image conspicuousness detection method based on motion color-associations
CN102867188B (en) Method for detecting seat state in meeting place based on cascade structure
CN111611874B (en) Face mask wearing detection method based on ResNet and Canny
CN104166841A (en) Rapid detection identification method for specified pedestrian or vehicle in video monitoring network
CN103942525A (en) Real-time face optimal selection method based on video sequence
CN105740945A (en) People counting method based on video analysis
CN102609724B (en) Method for prompting ambient environment information by using two cameras
CN113095263B (en) Training method and device for pedestrian re-recognition model under shielding and pedestrian re-recognition method and device under shielding
CN107369158A (en) The estimation of indoor scene layout and target area extracting method based on RGB D images
CN105069816B (en) A kind of method and system of inlet and outlet people flow rate statistical
CN102034267A (en) Three-dimensional reconstruction method of target based on attention
CN107944437B (en) A kind of Face detection method based on neural network and integral image
CN107123130A (en) Kernel correlation filtering target tracking method based on superpixel and hybrid hash
CN104537381B (en) A kind of fuzzy image recognition method based on fuzzy invariant features
CN107944403A (en) Pedestrian's attribute detection method and device in a kind of image
CN107832730A (en) Improve the method and face identification system of face recognition accuracy rate
CN108960142A (en) Pedestrian based on global characteristics loss function recognition methods again

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
GR01 Patent grant
GR01 Patent grant