CN106203428A - The image significance detection method merged based on blur estimation - Google Patents

Abstract

The present invention provides a kind of image significance detection method merged based on blur estimation, stage and fuzziness feature application stage is obtained including Low Level Vision feature, described Low Level Vision feature obtains the image that stage input is to be detected, utilize classical bottom-up significance detection algorithm that image is detected, and significant characteristics figure detection obtained is as the Low Level Vision feature acquisition result of image to be detected.Described fuzziness feature application stage, first input organize training image blocks more, train sparse dictionary；Mode followed by blur estimation is simulated psychological peculiarity when photographer shoots and quantifies it, quantized result is used for instructing the fusion of the significance information under different mechanisms, obtains final saliency testing result, to improve significance accuracy of detection.Saliency testing result obtained by the present invention more conforms to human vision significance detection pattern.Meanwhile, there is more preferable robustness and higher accuracy of detection.

Description

The image significance detection method merged based on blur estimation

Technical field

The present invention relates to saliency detection field, be specifically related to the detection of a kind of saliency based on blur estimation Method.

Background technology

In recent years, along with the development of digital technology, photo size is increasing, and resolution is more and more higher, photo institute The information contained also becomes increasingly to enrich, it is desirable to process the full detail in image in time for computer image system It it is a no small challenge.Select mechanism to assist the information in image to carry out selectivity process, to subtract it would therefore be desirable to have a kind of The burden of light image processing system.The vision significance of the mankind can help to contain in human visual system's resolution image important letter The region of breath, reduces the interference of background area simultaneously.By vision significance model is applied to Computer Image Processing field, Can more quickly extract the information in image, improve image procossing precision.Therefore, the Visual Observations Observations mistake of how simulating human Journey, it is thus achieved that more efficiently accurate image significance detection method, becomes the problem that computer vision field is urgently to be resolved hurrily.

The mode being presently used for saliency detection can be divided into two kinds from principle, and one is to utilize Low Level Vision Feature, such as color characteristic, the bottom-up significance detection mode of direction character and contrast metric etc..This detection side Formula data characteristics based on image, has detection rates fast, the multiple advantages such as computation complexity is low.But this traditional significance Detection mode easily produces detection mistake, thus is seldom used alone.Another kind is to utilize high-rise visual signature, i.e. utilizes priori Knowledge assists to carry out image the top-down significance detection mode of significance detection.This significance detection mode detects Precision is high, and can adjust flexibly along with different Detection tasks, but computation complexity is higher, the longest. The research that current significance detection field is the most popular is bottom-up significance detection mode aobvious with top-down Work property detection mode blends, and has both utilized the Low Level Vision feature of image also to utilize the high-rise visual signature of image, has drawn above-mentioned The advantage of two kinds of significance detection modes, it is thus achieved that more efficiently accurate significance testing result.But the fusion of current popular is certainly The end, is upwards not directed to psychological peculiarity when utilizing photographer to shoot and assists to carry out significantly with top-down significance detection mode Property detection, and the shooting psychology of photographer is significant for the understanding of image.Therefore, the present invention is directed to significantly Property the deficiency that exists of detection method, the creative psychology quantified when photographer shoots that proposes is intended to and as priori Auxiliary carries out saliency detection.

Summary of the invention

The present invention is directed to the deficiency that traditional images significance detection method exists, propose a kind of based on blur estimation fusion The technical scheme of image significance detection method.

Technical solution of the present invention provides a kind of image significance detection method merged based on blur estimation, regards including low layer Feel that feature obtains stage and fuzziness feature application stage；

The described Low Level Vision feature acquisition stage comprises the steps of,

Step 1.1, inputs test image to be detected,

Step 1.2, according to the test image of step 1.1 input, utilizes bottom-up saliency detection algorithm to obtain Take the Low Level Vision feature of image, generate the significant characteristics figure of Low Level Vision corresponding to image to be detected, use S_BRepresent；

The described fuzziness feature application stage comprises the steps of

Step 2.1, input is organized training image blocks more, is often organized the image block P that training image blocks includes focusing on_fWith the figure defocused As block P_d；

Step 2.2, for each training image group Y={y of input₁,...,y_n, carry out vectorization, y₁,...,y_nRepresent Concrete image block in one training image group, n represents the quantity of image block in a training image group；

And train sparse dictionary D to make to input data can be expressed as equation below,

$\begin{array}{cc}\underset{x_i}{min}\left|\right|y_i-{\mathrm{Dx}}_i|{|}_2^2& s.t.\left|\right|x_i|{|}_0\≤k\end{array}$

Wherein, x_iRepresent is to utilize different dictionary atom to blend fitted figure as y_iTime the most homoatomic weight parameter, K represents x_iThe value of sparse degree；

Step 2.3, for step 2.2 gained sparse dictionary D, the picture number to be detected of read step 1.1 input it is believed that Breath calculates, it is thus achieved that blur estimation figure S_D；

Step 2.4, plot step 2.3 gained blur estimation figure S_DCorresponding grey level histogram H_D, and according to intensity histogram Map analysis obtains the maximum I of blur estimation figure_maxWith minima I_min；

Step 2.5, utilizes step 2.3 gained blur estimation figure S_DMaximum I_maxWith minima I_minCalculate image Michelson contrast C, computing formula is as follows,

$C=\frac{I_{max}-I_{\min }}{I_{max}+I_{\min }}$

Step 2.6, Michelson contrast C calculated for step 2.5, utilize equation below to be mapped to default Within scope,

$\mathrm{\λ}=\frac{1}{1+e^{-(a+bC)}}$

Wherein, to be that photographer shoots the value of intent parameter, parameter a and parameter b corresponding to preset range for λ；

Step 2.7, according to step 2.3 gained blur estimation figure S_D, obtain the focal zone in image to be detected, with focusing The local maximum point in region is as vision localization point F_i, Gaussian smoothing is carried out for each vision localization point, obtains View-based access control model anchor point F_iVisual density figure D_i, by all vision localization point F_iCorresponding visual density figure D_iSuperposition obtains base In the significant characteristics figure that blur estimation merges, it is designated as the significant characteristics figure S of high-rise vision_T；

Step 2.8, utilizes step 2.6 gained photographer to shoot intent parameter λ, as merging weight, instructs difference The fusion of the lower significant characteristics figure of mechanism, fusion formula is as follows,

S=(1-λ) S_B+λS_T

Wherein, S represents the saliency testing result obtained, S_BFor the significant characteristics figure of Low Level Vision, S_TFor high level The significant characteristics figure of vision.

And, in step 2.6, preset range is [0.2,0.8].

And, the value of parameter a is-1.38, and the value of parameter b is 2.76.

The present invention compared with prior art has the advantage that

1. the present invention takes into full account the impact for image taking effect of mental activity when photographer shoots, the heart of photographing Of science introducing assists the significance carrying out image to detect, and more conforms to the perception of human visual system.

The shooting intent parameter introducing photographer of novelty the most of the present invention, and use it for assisting different significance information Fusion, there is more preferable theoretical basis, result is more accurate.

3. the significance detection method that the present invention proposes can be obviously improved the significance accuracy of detection of image, and has relatively Strong robustness, has certain dissemination.

Detailed description of the invention

The method that the present invention proposes is by bottom-up significance detection mode and top-down significance detection mode Blend, it is proposed that psychology when shooting photographer is simulated and the framework quantified, and simulation is used with the result quantified In the fusion of the significance information instructed under different mechanisms, the Low Level Vision feature of comprehensive utilization image is carried out with fuzziness feature Detection.

The present invention is first depending on the view data to be detected of input, calculates the Low Level Vision characteristic pattern of image；Secondly mould Intend psychological peculiarity when photographer shoots and carry out the blur estimation of image, calculate the vision localization point of viewer, obtain correspondence Visual density figure is as the high-rise visual signature figure of image.Quantify when photographer shoots followed by image blurring estimated result Intent parameter, strengthening photographer is intended by the object of (that is to say that viewer is easiest to be concerned about), utilizes quantized result to instruct High-rise visual signature figure and the fusion of Low Level Vision characteristic pattern, obtain final saliency testing result.The present invention simulates Psychological peculiarity when photographer shoots, meets the internal relation that image taking is intended between viewer's perception, obtained figure As significance testing result more conforms to human vision significance detection pattern.Meanwhile, the Low Level Vision of image is fully utilized Feature and high-rise visual signature, have more preferable robustness and higher accuracy of detection.

Technical solution of the present invention can use computer software mode to support automatic operational process, detailed below in conjunction with embodiment Technical solution of the present invention is described.

Embodiment includes that Low Level Vision feature obtains stage and fuzziness feature application stage.

The described Low Level Vision feature acquisition stage comprises the steps of

Step 1.1, inputs test image to be detected；

Step 1.2, according to the test image of step 1.1 input, utilizes classical bottom-up saliency detection Algorithm obtains the Low Level Vision feature of image, generates the Low Level Vision significant characteristics figure that image to be detected is corresponding, and uses S_BTable Show.

Obtain the stage by above Low Level Vision feature, test image to be detected in the present embodiment is all located Reason, obtains corresponding Low Level Vision significant characteristics figure, the fusion of significant characteristics under follow-up different mechanisms.

Classical bottom-up saliency detection algorithm can be found in documents below, and it will not go into details for the present invention:

[1]X.Hou and L.Zhang,“Saliency detection:a spectral residual approach,”in Proc.CVPR,2007.

[2]R.Achanta,S.Hemami,F.Estrada,and S.Susstrunk,“Frequency-tuned salient region detection,”in Proc.CVPR,2009.

The described fuzziness feature application stage comprises the steps of

Step 2.1, it is contemplated that photographer, during shooting image, strengthens often through change focal length and wanted to clap The object taken the photograph, so, the object more focused on more is likely to be photographer and wants the content of strengthening.Therefore, this method considers to utilize The mode of the blur estimation of image is assessed photographer in image and is wanted the content of strengthening.Input organizes different training images Block, often group training image blocks includes the image block P focused on_fWith the image block P defocused_d；Embodiment uses the image of a size of 8 × 8 Block is trained.

Described training image blocks derives to be split the random thousands of image captured, and carries out image block A certain degree of Gaussian Blur processes.Can be pre-selected voluntarily by those skilled in the art when being embodied as focusing image block and The image block defocused.

Step 2.2, for each training image group Y={y of input₁,…,y_n, carry out vectorization, y₁,…,y_nRepresent one Concrete image block in individual training image group, n represents the quantity of image block in a training image group, in each training image group Including the image block focused on and the image block defocused.

And train sparse dictionary D that input data can be expressed as equation below.

$\begin{array}{cc}\underset{x_i}{min}\left|\right|y_i-{\mathrm{Dx}}_i|{|}_2^2& s.t.\left|\right|x_i|{|}_0\≤k\end{array}$

Wherein, x_iRepresent is to utilize different dictionary atom to blend fitted figure as y_iTime the most homoatomic weight parameter, K is for representing x_iThe value of sparse degree.Formula can help fitting result more approaching to reality view data so that image blurring estimates Meter result is more accurate.

Step 2.3, for obtaining focus de-focus sparse dictionary D, the picture number to be detected of read step 1.1 input it is believed that Breath calculates, it is thus achieved that blur estimation figure S_D。

At blur estimation figure S_DIn, the gray value that each pixel is corresponding constitutes the dictionary atomic number table of this pixel Show.The dictionary atomic number constituted is the most, represents that the object corresponding to this pixel is the most clear, and the dictionary atomic number of composition is the fewest, The object representing corresponding is the fuzzyyest.

Calculating acquisition blur estimation figure and can be found in documents below, it will not go into details for the present invention:

[3]J.Shi,L.Xu,and J.Jia,“Just noticeable defocus blur detection and estimation,”in Proc.CVPR,2015.

Step 2.4, draws blur estimation figure S_DCorresponding grey level histogram H_D, and obtain according to intensity histogram map analysis The maximum I of blur estimation figure_maxWith minima I_min。

Grey level histogram H_DIn, gray scale the least expression picture material is the fuzzyyest, and gray scale the biggest expression picture material is the most clear. Same object is often made up of one group of pixel with close fog-level, and different objects is due to camera lens Distance difference and often there is different fog-levels.Therefore, grey level histogram presents different crests and trough.Ripple Peak represents the set of the pixel with close fog-level, thus corresponding same class object, trough then represents different objects Between fuzziness difference, may be used for distinguishing different objects.In the method, near the first of grey level histogram initial point The peak value that individual crest is corresponding is considered to represent average background pixel fuzziness, is recorded as I_min, and farthest from grey level histogram initial point Peak value corresponding to crest be considered in representative image the mean pixel fuzziness of the object focused on most, be recorded as I_max。

Step 2.5, utilizes blur estimation figure S_DMaximum I_maxWith minima I_minCalculate the Michelson contrast of image Degree C (visibility), computing formula is as follows:

$C=\frac{I_{max}-I_{min}}{I_{max}+I_{min}}$

Step 2.6, Michelson contrast C calculated for step 2.5, in the present embodiment by its numerical value profit It is mapped within the scope of [0.2,0.8] by equation below.When being embodied as, mapping range can be according to view data feature feelings Condition adjusts accordingly, and those skilled in the art can preset voluntarily, should control within the scope of [0,1].

$\mathrm{\λ}=\frac{1}{1+e^{-(a+bC)}}$

For calculated numerical value λ, it is defined as the shooting intent parameter of photographer, is used for describing photographer's shooting Time photography be intended in the picture represent situation.In the present embodiment, in order to Michelson contrast C value is controlled [0.2, 0.8], within, the value of parameter a is-1.38, and the value of parameter b is 2.76, and when being embodied as, those skilled in the art can be voluntarily The value of parameter preset a, b.

Step 2.7, calculates the blur estimation figure S obtained according to step 2.3_D, threshold value can be set when being embodied as and obtain Focal zone in image to be detected, in the present embodiment, threshold value value is 35, and the gray value region more than 35 is identified as treating The focal zone of detection image.With blur estimation figure S_DThe local maximum point of middle focal zone represents what this region focused on the most Local (the vision localization point being the most most possibly concerned about by the visual system of viewer), is recorded as F_i, for each vision Anchor point, carries out Gaussian smoothing to it, obtains based on this vision localization point F_iVisual density figure (Density Map), It is recorded as D_i, by all vision localization point F_iCorresponding visual density figure D_iSuperposition obtains the significance merged based on blur estimation Characteristic pattern, is recorded as S_T。

Step 2.8, the photographer utilizing step 2.6 to obtain shoots intent parameter λ, as merging weight, instructs not With the fusion of (Low Level Vision feature and fuzziness feature) significant characteristics figure under mechanism, fusion formula is as follows:

S=(1-λ) S_B+λS_T

Wherein, S represents the saliency testing result obtained, S_BFor Low Level Vision significant characteristics figure, S_TRegard for high level Feel significant characteristics figure, λ is the shooting intent parameter of photographer.

Specific embodiment described herein is only for example explanation spiritual to the present invention, technology neck belonging to the present invention Described specific embodiment can be made various amendment or supplements or use similar mode to substitute by the technical staff in territory, But do not deviate by the spirit of the present invention or surmount scope defined in appended claims.

Claims (3)

1. the image significance detection method merged based on blur estimation, it is characterised in that: include that Low Level Vision feature obtains Take stage and fuzziness feature application stage；

The described Low Level Vision feature acquisition stage comprises the steps of,

Step 1.1, inputs test image to be detected,

Step 1.2, according to the test image of step 1.1 input, utilizes bottom-up saliency detection algorithm acquisition figure The Low Level Vision feature of picture, generates the significant characteristics figure of Low Level Vision corresponding to image to be detected, uses S_BRepresent；

The described fuzziness feature application stage comprises the steps of

Step 2.1, input is organized training image blocks more, is often organized the image block P that training image blocks includes focusing on_fWith the image block defocused P_d；

Step 2.2, for each training image group Y={y of input₁,…,y_n, carry out vectorization, y₁,…,y_nRepresent a training Concrete image block in image sets, n represents the quantity of image block in a training image group；

And train sparse dictionary D to make to input data can be expressed as equation below,

$\begin{array}{cc}\underset{x_i}{min}\left|\right|y_i-{\mathrm{Dx}}_i|{|}_2^2& s.t.\left|\right|x_i|{|}_0\≤k\end{array}$

Wherein, x_iRepresent is to utilize different dictionary atom to blend fitted figure as y_iTime the most homoatomic weight parameter, k table Show x_iThe value of sparse degree；

Step 2.3, for step 2.2 gained sparse dictionary D, the image data information to be detected of read step 1.1 input is entered Row calculates, it is thus achieved that blur estimation figure S_D；

Step 2.4, plot step 2.3 gained blur estimation figure S_DCorresponding grey level histogram H_D, and divide according to grey level histogram Analysis obtains the maximum I of blur estimation figure_maxWith minima I_min；

Step 2.5, utilizes step 2.3 gained blur estimation figure S_DMaximum I_maxWith minima I_minCalculate image Michelson contrast C, computing formula is as follows,

$C=\frac{I_{max}-I_{min}}{I_{max}+I_{min}}$

Step 2.6, Michelson contrast C calculated for step 2.5, utilize equation below to be mapped to preset range Within,

$\mathrm{\λ}=\frac{1}{1+e^{-(a+bC)}}$

Wherein, to be that photographer shoots the value of intent parameter, parameter a and parameter b corresponding to preset range for λ；

Step 2.7, according to step 2.3 gained blur estimation figure S_D, obtain the focal zone in image to be detected, use focal zone Local maximum point as vision localization point F_i, Gaussian smoothing is carried out for each vision localization point, obtain based on Vision localization point F_iVisual density figure D_i, by all vision localization point F_iCorresponding visual density figure D_iSuperposition obtains based on mould Stick with paste the significant characteristics figure of estimation fusion, be designated as the significant characteristics figure S of high-rise vision_T；

Step 2.8, utilizes step 2.6 gained photographer to shoot intent parameter λ, as merging weight, instructs different mechanisms The fusion of lower significant characteristics figure, fusion formula is as follows,

S=(1-λ) S_B+λS_T

Wherein, S represents the saliency testing result obtained, S_BFor the significant characteristics figure of Low Level Vision, S_TFor high-rise vision Significant characteristics figure.

The image significance detection method merged based on blur estimation the most according to claim 1, it is characterised in that: step In 2.6, preset range is [0.2,0.8].

The image significance detection method merged based on blur estimation the most according to claim 2, it is characterised in that: parameter a Value be-1.38, the value of parameter b is 2.76.