CN102722520A - Method for classifying pictures by significance based on support vector machine - Google Patents

Abstract

The invention relates to a method for classifying pictures by significance, wherein the pictures are on web pages whose content can be accessed without obstacles. The method is characterized in that part of web pages are captured from the internet to establish training data; that as to every picture in each web page, space features and structure features of the picture which can represent the picture formally are extracted, and a significance class label is marked artificially on the picture according to functions of the picture in the web page; that acquired picture feature data and corresponding class labels are trained to form a classifier through support vector machine algorithm; and that pending web pages are captured, space features and structure features of all pictures on the web page are extracted, and the pictures are classified into significant class and insignificant class by the trained classifier. The method is advantageous in that pictures which are significant to the visually disabled can be selected and that substitution texts can be selectively provided for the significant pictures, which is convenient for the visually disabled to acquire picture information on the web pages.

Description

A kind of picture classifying importance method based on SVMs

Technical field

The present invention relates to the technical field of accessible detection of webpage and remodeling method, particularly based on the picture classifying importance method of SVMs.

Background technology

According to whole nation disabled person's sample survey second time result in 2006, China had 1,233 ten thousand people with visual disabilities now, and they are colonies that particular difficulty goes up in society., along with the height of internet is popularized and the internet continuous rising of importance in daily life, the accessible access problem of people's with visual disabilities info web becomes the important content of the accessible building-up work of China's information.People with visual disabilities is because the health obstacle, and the content on can't the normal reading webpage is so its indulging in the internet is outstanding especially.People with visual disabilities uses usually and reads to shield the softward interview web page contents.Reading to shield software helps the disabled person and obtains information through the text message on the webpage being converted into voice.For the picture on the webpage, read to shield the alternative textual description that is comprised in ALT attribute, LONGDESC attribute or the picture header of software through the picture in the html document < IMG>label, come to describe image content to user with visual disabilities.Increasing picture on the existing webpage, promoted the twenty-twenty vision user surf the Net experience in, but give the people's with visual disabilities more obstacle of web page contents visit the interpolation.

Be not that all pictures all need substitute textual description in the webpage.Understanding to web page contents in the webpage is important with the relevant picture of navigation, and suitable alternative text need be provided.In addition, also have many picture right and wrong that are used to promote webpage visual effect or structure of web page important on the webpage, empty alternative text should be provided, otherwise can disturb people's with visual disabilities information to obtain.Therefore coming that through a kind of method the picture on the webpage is carried out classifying importance seems particularly important.

At present, in fields such as machine learning, the research of sorting algorithm is reached its maturity.Picture in the webpage is extracted characteristic; And the method that marks with manual work is ready to training dataset; Adopt the method for existing SVMs in the machine learning to obtain a sorter afterwards, just can picture be carried out classifying importance at last through the sorter that obtains from the SVMs training according to training data.

Summary of the invention

In order to distinguish important picture and the non-important picture on the webpage; Thereby only alternative text is provided for important picture; Make people with visual disabilities can obtain the information on the webpage better, the present invention proposes a kind of picture classifying importance method based on SVMs, this method may further comprise the steps:

1, the Web page picture classifying importance method of the accessible visit of a kind of object web page content, the step of this method is following:

1) grasps some webpages, the picture in the locating web-pages from the internet;

2) Web page picture that step 1) is obtained extracts space characteristics and architectural feature, to its importance classes label of the artificial mark of every pictures, all pictures is divided into important and non-important two types, obtains training data;

3) utilize algorithm of support vector machine, on training data, train a sorter;

4) grasp the webpage that will carry out picture classification from the internet; All pictures on the webpage extract its space characteristics and architectural feature equally; Use the sorter that obtains in the step 3), the importance of predicted pictures is divided into important and non-important two types to the picture on the webpage according to this.

2, the picture described in the step 1) comprises:

1) <img>element in the webpage;

2) < object>and < embed>element that has vision to appear;

3) comprise the background picture attribute, content is empty node.

3, the space characteristics step 2) comprises:

1) picture is with respect to the relative width nw of screen, and just the ratio of picture width and screen width when ratio surpasses 1, directly gets 1 as width normalization characteristic, and final computing formula is following:

$\mathrm{nw}=\min (\frac{\mathrm{iw}}{\mathrm{sw}},1)$

Wherein iw is the picture width, and sw is a screen width;

2) picture is with respect to the relative height nh of screen, and just the ratio of picture height and screen height when ratio surpasses 1, directly gets 1 as height normalization characteristic, and final computing formula is following:

$\mathrm{nh}=\min (\frac{\mathrm{ih}}{\mathrm{sh}},1)$

Wherein ih is the picture height, and sh is a screen height;

3) the lateral attitude nlo of picture in webpage is the horizontal relative position of center picture in screen, and computing formula is:

$\mathrm{nlo}=\min (\frac{\mathrm{lo}+\mathrm{iw}/2}{\mathrm{sw}},1)$

Wherein lo is the distance of the left frame of picture left frame and webpage;

4) the relative distance nto of picture and page top, computing formula is:

$\mathrm{nto}=\min (\frac{\mathrm{to}}{\mathrm{sh}},1)$

Wherein to is the distance of picture upper side frame and page top;

5) the relative distance nbo of the picture and page bottom, computing formula is:

$\mathrm{nbo}=\min (\frac{\mathrm{bo}}{\mathrm{sh}},1)$

Wherein bo is the distance of picture lower frame and page bottom.

4, the architectural feature step 2) comprises:

1) arranged side by side picture number n lc, i.e. the number of picture in the set of pictures of photo current place, a group of promptly existing in the webpage is identical with the photo current length and width, and with laterally, the number of the picture arranged of mode vertical or form; Need numerical value be normalized to [0,1] interval, computing formula is:

nlc＝1-0.5 ^lc-1

Wherein lc is the picture number of photo current place set of pictures;

2) the link quality characteristic of picture divides three kinds of situation to consider, if picture is a link; And in webpage, have the text link that points to same address, then this eigenwert is 0, if picture is not link; Then this eigenwert is 0.5; If picture is a link, and in same webpage, do not have the text link that points to same address, then this eigenwert is 1;

3) the picture heel is counted nwc with the speech of text, also is divided into two kinds of situation, if picture is a link, it is 0.5 that eigenwert then is set, otherwise calculates by following formula:

nwc＝1-0.5 ^wc

Wherein wc is the speech number of picture followed with text;

4) area of picture place webpage hyperlink content accounts for the proportion nna of page all the elements area, and computing formula is following:

$\mathrm{nna}=\frac{\mathrm{na}}{\mathrm{ca}}$

Wherein na is a web page interlinkage content area, and ca is the area of page all the elements.

5, the said algorithm of support vector machine training process of step 3) is following:

1) given training data { (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), x wherein _iBe the space characteristics of i pictures and the vector that architectural feature is formed, y _iBe the class label of i pictures, when promptly picture is important, y _iBe 1, otherwise y _iBe-1, parameters C of using in the selected algorithm and γ;

2) maximization is about α=(α ₁, α ₂..., α _n)

Function is tried to achieve

The value of α when reaching maximum, need use this value in the time of classification:

$\tilde{L}\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_{j}k(x_i,x_j)$

Guarantee following constraint condition establishment simultaneously:

0≤α ₁≤C

$\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i=0$

Wherein C is a parameter selected in the step 1), k (x _i, x _j) be kernel function, what adopt here is gaussian kernel function, expression formula is:

k(x _i，x _j)＝exp?(-γ||x _i-x _j|| ²)

γ is a parameter selected in the step 1), the radial effect scope of control gaussian kernel function;

3) calculate b:

$b=y_i-\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_j{y}_{j}k(x_i,x_j)$

B is the intercept on classification plane, can use the value of the b that obtains in the follow-up assorting process here.

6, the said sorter assorting process of step 4) is following:

1) utilize variable α and the b that tries to achieve in the training process, the proper vector x of every picture to be classified calculated following decision function f (x):

$f\left(x\right)=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_{i}k(x_i,x)+b)$

The sgn function is a sign function, and expression formula is following:

$\mathrm{sgn}\left(x\right)=\left\{\begin{array}{cc}1& x>0\\ 0& x=0\\ -1& x＜0\end{array}\right.$

2) when the value of decision function f (x) is+1, then corresponding picture is assigned to important one type, when the decision function value is-1, corresponding picture is assigned to non-important one type.

The present invention proposes picture classifying importance method based on SVMs; Its advantage is: can from webpage, filter out picture important concerning the visual disability people;, make things convenient for the visual disability people to obtain info web and lay the first stone optionally for important picture provides alternative text at the back.This method is applicable to all types of webpages, need not the backstage and classifies.

Description of drawings

Fig. 1 is a method flow diagram of the present invention.

Embodiment

1, the Web page picture classifying importance method of the accessible visit of a kind of object web page content, the step of this method is following:

1) grasps some webpages, the picture in the locating web-pages from the internet;

2) Web page picture that step 1) is obtained extracts space characteristics and architectural feature, to its importance classes label of the artificial mark of every pictures, all pictures is divided into important and non-important two types, obtains training data;

3) utilize algorithm of support vector machine, on training data, train a sorter;

4) grasp the webpage that will carry out picture classification from the internet; All pictures on the webpage extract its space characteristics and architectural feature equally; Use the sorter that obtains in the step 3), the importance of predicted pictures is divided into important and non-important two types to the picture on the webpage according to this.

2, the picture described in the step 1) comprises:

1) <img>element in the webpage;

2) < object>and < embed>element that has vision to appear;

3) comprise the background picture attribute, content is empty node.

3, the space characteristics step 2) comprises:

1) picture is with respect to the relative width nw of screen, and just the ratio of picture width and screen width when ratio surpasses 1, directly gets 1 as width normalization characteristic, and final computing formula is following:

$\mathrm{nw}=\min (\frac{\mathrm{iw}}{\mathrm{sw}},1)$

Wherein iw is the picture width, and sw is a screen width;

2) picture is with respect to the relative height nh of screen, and just the ratio of picture height and screen height when ratio surpasses 1, directly gets 1 as height normalization characteristic, and final computing formula is following:

$\mathrm{nh}=\min (\frac{\mathrm{ih}}{\mathrm{sh}},1)$

Wherein ih is the picture height, and sh is a screen height;

3) the lateral attitude nlo of picture in webpage is the horizontal relative position of center picture in screen, and computing formula is:

$\mathrm{nlo}=\min (\frac{\mathrm{lo}+\mathrm{iw}/2}{\mathrm{sw}},1)$

Wherein lo is the distance of the left frame of picture left frame and webpage;

4) the relative distance nto of picture and page top, computing formula is:

$\mathrm{nto}=\min (\frac{\mathrm{to}}{\mathrm{sh}},1)$

Wherein to is the distance of picture upper side frame and page top;

5) the relative distance nbo of the picture and page bottom, computing formula is:

$\mathrm{nbo}=\min (\frac{\mathrm{bo}}{\mathrm{sh}},1)$

Wherein bo is the distance of picture lower frame and page bottom.

4, the architectural feature step 2) comprises:

1) arranged side by side picture number n lc, i.e. the number of picture in the set of pictures of photo current place, a group of promptly existing in the webpage is identical with the photo current length and width, and with laterally, the number of the picture arranged of mode vertical or form; Need numerical value be normalized to [0,1] interval, computing formula is:

nlc＝1-0.5 ^lc-1

Wherein lc is the picture number of photo current place set of pictures;

2) the link quality characteristic of picture divides three kinds of situation to consider, if picture is a link; And in webpage, have the text link that points to same address, then this eigenwert is 0, if picture is not link; Then this eigenwert is 0.5; If picture is a link, and in same webpage, do not have the text link that points to same address, then this eigenwert is 1;

3) the picture heel is counted nwc with the speech of text, also is divided into two kinds of situation, if picture is a link, it is 0.5 that eigenwert then is set, otherwise calculates by following formula:

nwc＝1-0.5 ^wc

Wherein wc is the speech number of picture followed with text;

4) area of picture place webpage hyperlink content accounts for the proportion nna of page all the elements area, and computing formula is following:

$\mathrm{nna}=\frac{\mathrm{na}}{\mathrm{ca}}$

Wherein na is a web page interlinkage content area, and ca is the area of page all the elements.

5, the said algorithm of support vector machine training process of step 3) is following:

1) given training data { (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), x wherein _iBe the space characteristics of i pictures and the vector that architectural feature is formed, y _iBe the class label of i pictures, when promptly picture is important, y _iBe 1, otherwise y _iBe-1, parameters C of using in the selected algorithm and γ;

2) maximization is about α=(α ₁, α ₂..., α _n)

Function is tried to achieve

The value of α when reaching maximum, need use this value in the time of classification:

$\tilde{L}\left(\mathrm{\&alpha;}\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i-\frac{1}{2}\underset{\mathrm{ij}}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_j{y}_i{y}_{j}k(x_i,x_j)$

Guarantee following constraint condition establishment simultaneously:

0≤α ₁≤C

$\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_i=0$

Wherein C is a parameter selected in the step 1), k (x _i, x _j) be kernel function, what adopt here is gaussian kernel function, expression formula is:

k(x _i，x _j)＝exp(-γ||x _i-x _j|| ²)

γ is a parameter selected in the step 1), the radial effect scope of control gaussian kernel function;

3) calculate b:

$b=y_i-\underset{j}{\mathrm{\&Sigma;}}{\mathrm{\&alpha;}}_j{y}_{j}k(x_i,x_j)$

B is the intercept on classification plane, can use the value of the b that obtains in the follow-up assorting process here.

6, the said sorter assorting process of step 4) is following:

1) utilize variable α and the b that tries to achieve in the training process, the proper vector x of every picture to be classified calculated following decision function f (x):

$f\left(x\right)=\mathrm{sgn}(\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{y}_{i}k(x_i,x)+b)$

The sgn function is a sign function, and expression formula is following:

$\mathrm{sgn}\left(x\right)=\left\{\begin{array}{cc}1& x>0\\ 0& x=0\\ -1& x＜0\end{array}\right.$

2) when the value of decision function f (x) is+1, then corresponding picture is assigned to important one type, when the decision function value is-1, corresponding picture is assigned to non-important one type.

The described content of this instructions embodiment only is enumerating the way of realization of inventive concept; Should not being regarded as of protection scope of the present invention only limits to the concrete form that embodiment states, protection scope of the present invention also reach in those skilled in the art conceive according to the present invention the equivalent technologies means that can expect.

Claims (6)

1. the Web page picture classifying importance method of the accessible visit of an object web page content the method is characterized in that:

1) grasps webpage, the picture in the locating web-pages from the internet;

2) Web page picture that step 1) is obtained extracts space characteristics and architectural feature, to its importance classes label of the artificial mark of every pictures, all pictures is divided into important and non-important two types, obtains training data;

3) utilize algorithm of support vector machine, on training data, train a sorter;

4) grasp the webpage that will carry out picture classification from the internet; All pictures on the webpage extract its space characteristics and architectural feature equally; Use the sorter that obtains in the step 3), the importance of predicted pictures is divided into important and non-important two types to the picture on the webpage according to this.

2. the Web page picture classifying importance method of the accessible visit of object web page content as claimed in claim 1, it is characterized in that: the picture described in the described step 1) comprises:

1) <img>element in the webpage;

2) < object>and < embed>element that has vision to appear;

3) comprise the background picture attribute, content is empty node.

3. the Web page picture classifying importance method of the accessible visit of object web page content as claimed in claim 1, it is characterized in that: the space characteristics described step 2) comprises:

1) picture is with respect to the relative width nw of screen, and just the ratio of picture width and screen width when ratio surpasses 1, directly gets 1 as width normalization characteristic, and final computing formula is following:

Wherein iw is the picture width, and sw is a screen width;

2) picture is with respect to the relative height nh of screen, and just the ratio of picture height and screen height when ratio surpasses 1, directly gets 1 as height normalization characteristic, and final computing formula is following:

Wherein ih is the picture height, and sh is a screen height;

3) the lateral attitude nlo of picture in webpage is the horizontal relative position of center picture in screen, and computing formula is:

Wherein lo is the distance of the left frame of picture left frame and webpage;

4) the relative distance nto of picture and page top, computing formula is:

Wherein to is the distance of picture upper side frame and page top;

5) the relative distance nbo of the picture and page bottom, computing formula is:

Wherein bo is the distance of picture lower frame and page bottom.

4. the picture classifying importance method of the accessible visit of object web page content as claimed in claim 1, it is characterized in that: the architectural feature described step 2) comprises:

1) arranged side by side picture number n lc, i.e. the number of picture in the set of pictures of photo current place, a group of promptly existing in the webpage is identical with the photo current length and width, and with laterally, the number of the picture arranged of mode vertical or form; Need numerical value be normalized to [0,1] interval, computing formula is:

nlc＝1-0.5 ^lc-1

Wherein lc is the picture number of photo current place set of pictures;

2) the link quality characteristic of picture divides three kinds of situation to consider, if picture is a link; And in webpage, have the text link that points to same address, then this eigenwert is 0, if picture is not link; Then this eigenwert is 0.5; If picture is a link, and in same webpage, do not have the text link that points to same address, then this eigenwert is 1;

3) the picture heel is counted nwc with the speech of text, also is divided into two kinds of situation, if picture is a link, it is 0.5 that eigenwert then is set, otherwise calculates by following formula:

nwc＝1-0.5 ^wc

Wherein wc is the speech number of picture followed with text;

4) area of picture place webpage hyperlink content accounts for the proportion nna of page all the elements area, and computing formula is following:

Wherein na is a web page interlinkage content area, and ca is the area of page all the elements.

5. the picture classifying importance method of the accessible visit of object web page content as claimed in claim 1 is characterized in that: the said algorithm of support vector machine training process of step 3) is following:

1) given training data { (x ₁, y ₁), (x ₂, y ₂) ..., (x _n, y _n), x wherein _iBe the space characteristics of i pictures and the vector that architectural feature is formed, y _iBe the class label of i pictures, when promptly picture is important, y _iBe 1, otherwise y _iBe-1, parameters C of using in the selected algorithm and γ;

2) maximization is about α=(α ₁, α ₂..., α _n)

Function is tried to achieve

The value of α when reaching maximum, need use this value in the time of classification:

Guarantee following constraint condition establishment simultaneously:

0≤α ₁≤C

Wherein C is a parameter selected in the step 1), k (x _i, x _j) be kernel function, what adopt here is gaussian kernel function, expression formula is:

k(x _i，x _j)＝exp(-γ||x _i-x _j|| ²)

γ is a parameter selected in the step 1), the radial effect scope of control gaussian kernel function;

3) calculate b:

B is the intercept on classification plane, can use the value of the b that obtains in the follow-up assorting process here.

6. the picture classifying importance method of the accessible visit of object web page content as claimed in claim 1 is characterized in that: the said sorter assorting process of step 4) is following:

1) utilize variable α and the b that tries to achieve in the training process, the proper vector x of every picture to be classified calculated following decision function f (x):

The sgn function is a sign function, and expression formula is following:

2) when the value of decision function f (x) is+1, then corresponding picture is assigned to important one type, when the decision function value is-1, corresponding picture is assigned to non-important one type.

Images