CN106227818A

CN106227818A - A kind of abnormal image data process and search method

Info

Publication number: CN106227818A
Application number: CN201610586482.2A
Authority: CN
Inventors: 黎海纤
Original assignee: Individual
Current assignee: Individual
Priority date: 2016-07-22
Filing date: 2016-07-22
Publication date: 2016-12-14

Abstract

The invention discloses a kind of abnormal image data to process and search method, it is characterized in that, processing and image retrieval including abnormal image data, abnormal image data processing method comprises the following steps: (1) abnormal image detection device generates an abnormal image generation message incoming abnormal data collection and treatment device when abnormal image being detected；(2) abnormal data collection and treatment device is according to the mark of abnormal image, generates abnormal image example, according to abnormal image instance properties information and abnormal Producing reason, search measure definition storehouse device, is found for the measure definition information of this exception.

Description

A kind of abnormal image data process and search method

Technical field

The present invention relates to field of image search, be specifically related to a kind of abnormal image data and process and search method.

Background technology

Along with the rise of the technology such as mobile Internet, Internet of Things, in global range, data volume rapidly increases, big data age Oneself is through arriving.Along with the generation of big data, it plays extremely important effect in modern society and economic activity.

When processing view data, many abnormal images can be run into, how it be processed and retrieval is a great problem.

Large-scale data gives traditional multimedia research, and the application and the research that are based especially on image bring new challenge And opportunity.

Summary of the invention

For the problems referred to above, the present invention provides a kind of abnormal image data to process and search method.

The purpose of the present invention realizes by the following technical solutions:

A kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data and image is examined Rope, abnormal image data processing method comprises the following steps:

(1) abnormal image detection device generates an abnormal image incoming exception of generation message when abnormal image being detected Data collection process device；

(2) abnormal data collection and treatment device is according to the mark of abnormal image, generates abnormal image example, according to Abnormal Map As instance properties information and abnormal Producing reason, search measure definition storehouse device, it is found for the measure definition letter of this exception Breath.

Preferably, it is characterized in that, abnormal image data processing method is further comprising the steps of: measure definition storehouse uses oneself Abnormal image feature is analyzed by learning functionality, then in measure definition storehouse, automatically generates the measure for this type of exception.

Preferably, it is characterized in that, abnormal image data processing method is further comprising the steps of: measure performs supplement device and holds Measure produced by row abnormal data collection and treatment device.

The invention have the benefit that the utilization by key word and labelling, in advance data base is divided into multiple subnumber According to storehouse, first retrieve in the subdata base that degree of association is high during retrieval, decrease amount of calculation, improve arithmetic speed；Based on When word bag represents image, it is proposed that weighting represents and First look similarity, decreases time overhead；Feature based combination table Diagram as time, make use of the space inclusion relation between local feature, propose to combine relevant local feature to increase Its visual expression ability strong；This feature combination not only has good yardstick and rotational invariance, but also can natural land productivity Carry out local geometric verification with the relative position information between each characteristic element, reject erroneous matching that may be present, to improve The accuracy rate of image retrieval；Characteristic quantification and combinational expression have been significantly increased the precision of retrieval.

Accompanying drawing explanation

The invention will be further described to utilize accompanying drawing, but the embodiment in accompanying drawing does not constitute any limit to the present invention System, for those of ordinary skill in the art, on the premise of not paying creative work, it is also possible to obtain according to the following drawings Other accompanying drawing.

Fig. 1 is that a kind of abnormal image data process and the schematic diagram of search method.

Detailed description of the invention

The invention will be further described with the following Examples.

Embodiment 1: a kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data And image retrieval, abnormal image data processing method comprises the following steps:

Preferably, it is characterized in that, image retrieval is realized by following steps:

(1) obtained information data by Cloud Server, after image information and non-image information are classified, be stored into respectively In image data base and non-picture data storehouse, and the identical image information in source and non-image information are marked；

(2) non-picture data storehouse is divided into multiple first subdata base according to multiple key words set in advance, simultaneously According to described key word, non-image information is classified, be stored into respectively in described first subdata base of correspondence, do not mate Non-image information to key word is individually divided into a class；

(3) image data base is divided into multiple second subdata base according to described key word, and to image information according to Identical non-image information of originating with it is classified, and is stored into respectively in corresponding described second subdata base, identical without source The image information of labelling, or the non-image information of its correspondence do not matches the image information of key word and is individually divided into a class；

(4) when user input key word retrieve time, according to the phase of key word and the key word set in advance of input Guan Du sequence is retrieved successively in each first numerical data base and the second subdata base respectively, and exports view data respectively Retrieval result with non-picture data；The image retrieval of each second subdata base additionally provides by input inquiry image I_p Retrieving the function of similar image, this function is realized by following steps:

(1-1) use SIFT feature that image local area is described；

(1 2) graphical representation based on word bag:

A. use and based on word bag model, local feature is quantified, if the sample being made up of M local feature vectors Space X={ x₁,…,x_M, to generate initial visual code book without monitor mode quick clustering, use arest neighbors strategy to set up local Feature x_jWith the mapping relations of corresponding vision word, j=1 ..., M；

B. set in codebook space and comprise N number of vision word, then any image is expressed as higher-dimension sparse vector { w₁,…,w_N, w_iThe weights of expression vision word i, i=1 ..., N；

w_{i} = \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}

In formula,Represent the number of times that in this image, vision word i occurs, t_iRepresent that in image data base, vision word i goes out Existing total degree, f_iRepresent the picture number comprising vision word i in image data base；

To weight w_iIt is normalized, order∑_iδ_i=1, the now higher-dimension sparse vector table of any image It is shown as { δ₁,…,δ_N}；

C. for any two width image p and q, define First look similarity S between two width images (p, q):

S (p, q) = \frac{1}{\sqrt{Σ_{i = 1}^{N} {(\frac{\frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}} - \frac{\frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}})}^{2}}}

S (p, q) the biggest, show two width images closer to.

Preferably, it is characterized in that, step (4) also includes:

The graphical representation of (1 3) feature based combination:

A, by comprise a main feature with large scale and by this main feature space cover several have relatively The set of the elemental characteristic of little yardstick is defined as feature combination, given any image P, extracts its feature combination of sets comprisedWherein, C_l=(Z_l,Y_l)；

In formula, C_lRepresent the l feature combination, Z_lIt is the main feature of the l feature combination, Y_lIt is the l feature combination One group element feature, Z_lAnd Y_lMeet 0.02 × a (P) ＜ a (Z_k),a(Y_l) ＜ 0.2 × a (P), and Y_lCorresponding local space district Territory is completely by main feature C_lLocal space cover, a () represent cover area of space area；

B, for any two feature combine C_f(Z_f,Y_f) and C_g(Z_g,Y_g), define the second vision phase of two feature combinations Seemingly spend R:

As cos (Z_f, Z_g) ＞ T₁Time:

R(C_f,C_g)=1{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

Work as T₁≥cos(Z_f, Z_g) ＞ T₂Time:

R(C_f,C_g)=0.5{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

As cos (Z_f, Z_g)≤T₂Time:

R(C_f,C_g)=0

In formula, T₁=(0.8,1), T₁=(0.5,0.8], cos () represents the cosine similarity of two features, Y_fi It is Y_fIn element, Y_giIt is Y_gIn element, { (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁Represent Y_fAnd Y_gIn The elemental characteristic number of coupling；

(1 4) image retrieval: for given query image I_p, first extract its feature combination of sets, by each feature Main Feature Mapping, to vision word, finds out database images I comprising this vision word_q, calculate it similar to query image Degree distance, by by similarity distance with setting threshold ratio relatively, completing image retrieval；

Described similarity distance uses below equation to calculate:

D=S (I_p,I_q)×maxR(I_p,I_q)

Wherein, (p q) represents the First look measuring similarity of two width images, maxR (I to S_p,I_q) represent two width images The combination of all features carry out the maximum of the second vision similarity tolerance.

A kind of abnormal image data at the present embodiment process and in search method, by key word and the utilization of labelling, In advance data base is divided into multiple subdata base, first retrieves in the subdata base that degree of association is high during retrieval, decrease Amount of calculation, improves arithmetic speed；When representing image based on word bag, it is proposed that weighting represents and First look similarity, subtracts Lack time overhead；Feature based combination table diagram as time, make use of the space inclusion relation between local feature, propose phase The local feature closed is combined to strengthen its visual expression ability；This feature combination not only has good yardstick and rotation Invariance, but also can naturally utilize the relative position information between each characteristic element to carry out local geometric verification, reject Erroneous matching that may be present, to improve the accuracy rate of image retrieval；Characteristic quantification and combinational expression have been significantly increased inspection The precision of rope；T₁=0.95, T₂=0.78, retrieval precision improves 50%, and retrieval rate improves 1%.

Embodiment 2: a kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data And image retrieval, abnormal image data processing method comprises the following steps:

(1-1) use SIFT feature that image local area is described；

(1 2) graphical representation based on word bag:

A. use and based on word bag model, local feature is quantified, if the sample being made up of M local feature vectors Space X={ x₁,…,x_M, to generate initial visual code book without monitor mode quick clustering, use arest neighbors strategy to set up local Feature x_lWith the mapping relations of corresponding vision word, j=1 ..., M；

w_{i} = \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}

S (p, q) = \frac{1}{\sqrt{Σ_{i = 1}^{N} {(\frac{\frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}} - \frac{\frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}})}^{2}}}

S (p, q) the biggest, show two width images closer to.

Preferably, it is characterized in that, step (4) also includes:

The graphical representation of (1 3) feature based combination:

In formula, C_lRepresent the l feature combination, Z_lIt is the main feature of the l feature combination, Y_lIt is the l feature combination One group element feature, Z_lAnd Y_lMeet 0.02 × a (P) ＜ a (Z_l),a(Y_l) ＜ 0.2 × a (P), and Y_lCorresponding local space district Territory is completely by main feature C_lLocal space cover, a () represent cover area of space area；

As cos (Z_f, Z_g) ＞ T₁Time:

R(C_f,C_g)=1{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

Work as T₁≥cos(Z_f, Z_g) ＞ T₂Time:

R(C_f,C_g)=0.5{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＜ Y₁}

As cos (Z_f, Z_g)≤T₂Time:

R(C_f,C_g)=0

Described similarity distance uses below equation to calculate:

D=S (I_p,I_q)×maxR(I_p,I_q)

Embodiment 3: a kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data And image retrieval, abnormal image data processing method comprises the following steps:

(1-1) use SIFT feature that image local area is described；

(1 2) graphical representation based on word bag:

w_{i} = \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}

S (p, q) = \frac{1}{\sqrt{Σ_{i = 1}^{N} {(\frac{\frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}} - \frac{\frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}})}^{2}}}

S (p, q) the biggest, show two width images closer to.

Preferably, it is characterized in that, step (4) also includes:

The graphical representation of (1 3) feature based combination:

As cos (Z_f, Z_g) ＞ T₁Time:

R(C_f,C_g)=1{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

Work as T₁≥cos(Z_f, Z_g) ＞ T₂Time:

R(C_f,C_g)=0.5{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

As cos (Z_f, Z_g)≤T₂Time:

R(C_f,C_g)=0

Described similarity distance uses below equation to calculate:

D=S (I_p,I_q)×maxR(I_p,I_q)

Embodiment 4: a kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data And image retrieval, abnormal image data processing method comprises the following steps:

(1-1) use SIFT feature that image local area is described；

(1 2) graphical representation based on word bag:

w_{i} = \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}

S (p, q) = \frac{1}{\sqrt{Σ_{i = 1}^{N} {(\frac{\frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}} - \frac{\frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}})}^{2}}}

S (p, q) the biggest, show two width images closer to.

Preferably, it is characterized in that, step (4) also includes:

The graphical representation of (1 3) feature based combination:

As cos (Z_f, Z_g) ＞ T₁Time:

R(C_f,C_g)=1{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

Work as T₁≥cos(Z_f, Z_g) ＞ T₂Time:

R(C_f,C_g)=0.5{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

As cos (Z_f, Z_g)≤T₂Time:

R(C_f,C_g)=0

Described similarity distance uses below equation to calculate:

D=S (I_p,I_q)×maxR(I_p,I_q)

Embodiment 5: a kind of abnormal image data process and search method, it is characterized in that, process including abnormal image data And image retrieval, abnormal image data processing method comprises the following steps:

(1-1) use SIFT feature that image local area is described；

(1 2) graphical representation based on word bag:

B. set in codebook space and comprise N number of vision word, then any image is expressed as higher-dimension sparse vector { w₁..., w_N, w_iThe weights of expression vision word i, i=1 ..., N；

w_{i} = \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}

S (p, q) = \frac{1}{\sqrt{Σ_{i = 1}^{N} {(\frac{\frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{p}}{t_{i} - t_{i}^{p}} \times \frac{1}{f_{i}}} - \frac{\frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}}{Σ_{i} \frac{t_{i}^{q}}{t_{i} - t_{i}^{q}} \times \frac{1}{f_{i}}})}^{2}}}

S (p, q) the biggest, show two width images closer to.

Preferably, it is characterized in that, step (4) also includes:

The graphical representation of (1 3) feature based combination:

As cos (Z_f, Z_g) ＞ T₁Time:

R(C_f,C_g)=1{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

Work as T₁≥cos(Z_f, Z_g) ＞ T₂Time:

R(C_f,C_g)=0.5{ (Y_fi,Y_gj)|Y_fi∈Y_f,Y_gj∈Y_g,cos(Y_f, Y_g) ＞ T₁}

As cos (Z_f, Z_g)≤T₂Time:

R(C_f,C_g)=0

Described similarity distance uses below equation to calculate:

D=S (I_p,I_q)×maxR(I_p,I_q)

Last it should be noted that, above example is only in order to illustrate technical scheme, rather than the present invention is protected Protecting the restriction of scope, although having made to explain to the present invention with reference to preferred embodiment, those of ordinary skill in the art should Work as understanding, technical scheme can be modified or equivalent, without deviating from the reality of technical solution of the present invention Matter and scope.

Claims

1. abnormal image data process and a search method, it is characterized in that, process and image retrieval including abnormal image data, Abnormal image data processing method comprises the following steps:

(1) abnormal image detection device generates an abnormal image incoming abnormal data of generation message when abnormal image being detected Collection and treatment device；

(2) abnormal data collection and treatment device is according to the mark of abnormal image, generates abnormal image example, real according to abnormal image Example attribute information and abnormal Producing reason, search measure definition storehouse device, it is found for the measure definition information of this exception.

A kind of abnormal image data the most according to claim 1 process and search method, it is characterized in that, abnormal image data Processing method is further comprising the steps of: measure definition storehouse uses self-teaching function to be analyzed abnormal image feature, then In measure definition storehouse, automatically generate the measure for this type of exception.

A kind of abnormal image data the most according to claim 2 process and search method, it is characterized in that, abnormal image data Processing method is further comprising the steps of: measure performs supplement device and performs measure produced by abnormal data collection and treatment device.