CN112463804B - A Data Processing Method of Image Database Based on KDTree - Google Patents

Abstract

The invention discloses a KDTree-based image database data processing method, comprising the following steps: step 1, traversing and integrating map labeling information based on KDTree to obtain label set S={s1,s2,...,sn}; step 2 Step 3: Perform corresponding desensitization processing according to the sensitivity level of the map annotation content. The present invention utilizes the location information marked on the map, realizes the traversal and integration of the marked content in the geospatial data and the automatic processing of sensitive information desensitization, and overcomes the cumbersome work, low efficiency, error-prone and loopholes under the existing manual processing. The phenomenon.

Description

A Data Processing Method of Image Database Based on KDTree

technical field

The invention relates to the field of geographic information data processing, in particular to a KDTree-based image database data processing method.

Background technique

In recent years, the rapid development of computer technology has promoted the development and progress of geographic information system technology. Geographic Information System (GIS) involves many disciplines such as geography, surveying and mapping, computer science and technology. It uses computers as tools, takes geospatial data as the research object, and combines the unique visualization effect of maps with geographic analysis. Functionality is integrated with database operations to provide decision-making information for many industries and sectors such as geography, planning and management.

At present, with the development of the mobile Internet, people's demand for services based on geographic information in their daily travel is increasing. The wide application of various electronic maps not only brings convenience to people's work and life, but also brings many hidden safety problems. Among them, the security protection of map annotation information is a problem worthy of study. The map annotation information may contain some sensitive content, such as national strategic resources, military forbidden areas, and national defense facilities. In this regard, the state has promulgated a number of legal regulations and policies, such as the Supplementary Provisions on the Representation of Public Map Content (Trial), and the Provisions on the Public Representation of Basic Geographic Information (Trial), etc., which clearly stipulate that the representation and The content that cannot be expressed has strengthened the security protection of geographic information from the legal and policy level. At present, various institutions have adopted the technical means of intranet isolation to ensure data security, and require that sensitive data in the intranet need to be desensitized before publishing geographic information to the extranet.

In the real scene, the labeling content on the map is saved in the form of an attribute table. In order to make the expression effect beautiful and complete, some labels are composed of multiple single characters. For example, in the attribute table, multiple records composed of single characters are used to represent a place name. When searching for keywords on the map, it is easy to miss some annotations. In this case, the existing desensitization work has to rely on manually checking the labeled content of each area in the map, and then review, identify and process it. This method still has the problem of content omission, and the map area above the city level It is huge, complex in content, cumbersome in manual processing and low in efficiency. Therefore, there is an urgent need to study a method for integrating and desensitizing the annotation information in the map based on computer technology, which is of great significance for maintaining the security of geographic information.

For the labeling content composed of multiple words, the common division method is lexical analysis, that is, converting the sequence of characters into a sequence of words, dividing the received string of consecutive characters into individual words, and then combining the obtained words with the sensitive vocabulary. Matching is performed to detect sensitive information. However, when performing routine addition, deletion or modification operations on multiple words included in the annotation, these words may be out of order and repeated in the attribute table, which cannot be handled by simple word segmentation.

Although the labeling content is irregularly arranged in the attribute table, each character in the map is associated with a location coordinate, and the labeling fields to be integrated have a strong correlation in location distribution, such as closer arrangement, top-to-bottom , from left to right, almost in a straight line, etc. The invention utilizes the location information marked on the map, and realizes the traversal, integration and desensitization processing of the marked content in the geospatial data.

Glossary: KDTree (k-dimensional tree): It is a tree data structure that stores instance points in k-dimensional space for fast retrieval.

jieba word segmentation: a very popular Chinese open source word segmentation package, with high performance, accuracy, scalability and other characteristics, currently mainly supports python, other languages also have related versions.

word2vec: is a group of related models used to generate word vectors. These models are shallow, two-layer neural networks that are trained to reconstruct linguistic word texts. The network is represented by words and needs to guess the input words in adjacent positions. Under the assumption of the bag-of-words model in word2vec, the order of words is not important. After the training is completed, the word2vec model can be used to map each word to a vector, which can be used to represent the relationship between words and words, which is the hidden layer of the neural network.

SUMMARY OF THE INVENTION

The invention proposes an image database data processing method based on KDTree (k-dimensional tree). The present invention utilizes the location information marked on the map, realizes the traversal and integration of the marked content in the geospatial data and the automatic processing of sensitive information desensitization, and overcomes the cumbersome work, low efficiency, error-prone and loopholes under the existing manual processing. The phenomenon.

To achieve the above object, technical scheme of the present invention is as follows:

A KDTree-based image database data processing method, comprising the following steps:

Step 1. Traverse and integrate the map annotation information based on KDTree, and obtain an annotation set S={s ₁ ,s ₂ ,...,s _n } composed of n annotations; sn represents the nth annotation;

Step 2: Sensitive information detection based on word similarity is performed on the label set S, and sensitivity grading is performed on the map label content;

Step 3: Perform corresponding desensitization processing according to the sensitivity level of the content marked on the map.

A further improvement, the step 1 includes the following steps:

1.1 Extract the position coordinates and character content of all words on the map in the attribute table of the map to form a data set, and construct KDTree according to the two-dimensional coordinates of each word;

1.2 Arrange all words in descending order of the ordinate to get the initial queue Q; create a mark array vis[] to record whether each word in the queue Q has been processed, initialize it to 0, and traverse the queue Q until the queue Q is empty;

1.3 Process the words in turn according to the arrangement order of the words in the preliminary queue Q;

If the current word point p is not processed, that is, vis[p]=0, execute step 1.4, and set vis[p] to 1;

If the current single point p has been processed, that is, vis[p]=1, jump to the next point in the queue Q;

1.4 Find the point in the constructed KDTree whose distance from the current word point p is the threshold [0, ε], and obtain the set of neighbor nodes of the current word point p, where ε represents the parameter of the integration range, and take the corresponding point of the previous word point p. 1.5-2 times the word width; search for a single-word point q that satisfies the integration condition with the current single-word point p in the order of the distance point p from the nearest to the farthest in the neighbor node set, if it is successfully found, replace the current single-word point with q , and juxtapose vis[q] to 1; the single-character point q is the corresponding point of the single-character q in KDTree;

1.5 Repeat step 1.4 until there is no single-character point that can be integrated with the current single-character point in the set of adjacent nodes, then the integrated single-character point is used as a label;

1.6 When there is no single word point that can be integrated with the current single word point in the set of adjacent nodes, the next unprocessed single word is processed according to the arrangement order of the words in the preliminary queue Q;

1.7 Repeat steps 1.3-1.6 until the words in the preliminary queue Q are all processed, and each label on the map is obtained.

Further improvement, in step 1.4, the integration conditions are required as follows:

The single-character point q has not been processed, that is, vis[q]=0;

Situation 1. When the integrated field contains only one current single-character point p, the current single-character point p is integrated with the single-character point q that is closest to the previous single-character point p in the set of adjacent nodes; Then the single character point p itself constitutes a label;

Case 2. When the integrated field contains two or more characters, that is, when the field composed of multiple characters is integrated with the single-character point q, determine whether all the words in the new field s composed of the single-character point q and the integrated field are not. Whether the range R of the arrays that are on the same straight line and formed by the distance of every two adjacent words satisfies:

表示新字段s中第i个单字点与第j个单字点j的距离，

和

分别表示取最大值和最小值； γ为新字段s中单字宽度的0.2-0.5倍；Among them, Len represents the number of words contained in the new field s formed,

Indicates the distance between the i -th word point and the j-th word point j in the new field s,

and

Respectively represent the maximum and minimum values; γ is 0.2-0.5 times the width of the word in the new field s;

If both of the above two integration conditions are satisfied, the field composed of multiple characters is integrated with the single-character dot q, and if at least one of them is not satisfied, the field composed of multiple characters is not integrated with the single-character dot q.

For further improvement, before integration, first eliminate the interference of repeated words. If the word boxes corresponding to p and q intersect and the content of the words is the same, then p and q are repeated words. Delete q in the attribute table to achieve deduplication.

A further improvement, in step 1.7, for the labeling of the horizontal distribution, the words in the label are arranged in the order of the abscissa from small to large, and from left to right.

A further improvement, the step 2 includes the following steps:

2.1: Use Chinese word segmentation technology to segment the marked content:

For each tag content si of tag set S, use Chinese word segmentation technology and word vector construction technology to convert into multiple word vectors; get si={a1,a2,...,am}, a1...am are m obtained after division feature word;

2.2: Convert feature words and sensitive words into word vectors using word vector construction technology:

Use word2vec to convert the feature words into word vectors, and the converted word vector of the jth feature word aj is denoted as Aj; similarly, convert all sensitive words bk in the sensitive thesaurus into word vectors, denoted as Bk; The similarity of sensitive words is quantified as the similarity between the feature word vector and the sensitive word vector, that is, the cosine of the angle between the inner product spaces of the two vectors is taken as the similarity, and the value range is [0, 1], and the closer the similarity is 1 indicates that the two words are more similar;

表示Aj与Bk的相似度；

Represents the similarity between Aj and Bk;

2.3 Sensitivity calculation of feature words:

，

值越大，则敏感词的敏感程度越高；遍历敏感词库，对特征词aj，定义特征词aj最大敏感度

为In the sensitive thesaurus, each sensitive word ck corresponds to a sensitivity level

,

The larger the value, the higher the sensitivity of the sensitive word; traverse the sensitive word database, define the maximum sensitivity of the feature word aj to the feature word aj

for

表示敏感词库，计算每个敏感词向量与特征词向量的相似度与敏感级别的乘积，取其中的最大值表示特征词的最大敏感度；in,

Represents the sensitive word library, calculates the product of the similarity between each sensitive word vector and the feature word vector and the sensitivity level, and takes the maximum value to represent the maximum sensitivity of the feature word;

大于θ时，特征词aj才具有敏感性，否则不认为是敏感词，敏感度记为0，即特征词的敏感度为Set the threshold parameter θ, when

When it is greater than θ, the feature word aj has sensitivity, otherwise it is not considered as a sensitive word, and the sensitivity is recorded as 0, that is, the sensitivity of the feature word is

2.4 Sensitivity calculation of annotation content:

的敏感度

为：define the ith label

sensitivity

for:

包含m个特征词，j表示是标注

的第j个特征词；

为标注

的敏感度，即为标注

包含的m个特征词敏感度累加的和；得出特征词与标注内容的敏感度后，将标注内容按敏感度划分为高、中、低、非敏感4个级别。In the formula, label

Contains m feature words, j means it is a label

The jth feature word of ;

to label

The sensitivity of , that is, the labeling

The sum of the sensitivities of the m feature words included; after obtaining the sensitivity of the feature words and the labeled content, the labeled content is divided into four levels of high, medium, low, and insensitive according to the sensitivity.

A further improvement, the step 3 includes the following steps:

3.1 Build a whitelist of geo-labeled information, and add the non-sensitive data to the whitelist every time the non-sensitive data misidentified by the algorithm is manually discovered to improve the fault tolerance rate;

3.2 After the sensitive data is screened through the whitelist, take measures to directly delete the marked content of high sensitivity level; extract the sensitive feature words for the marked content of medium and low sensitivity level, and randomly select the desensitization of deletion, replacement and generalization Then, the desensitized label sensitivity is recalculated. If it still does not meet the disclosure requirements after iterating for a preset number of times, the corresponding label will be completely deleted; among them, when selecting replacement, select the current sensitive feature word in the non-sensitive word database. The non-sensitive words with the highest similarity are used as replacements, and the specific content of the label description is abstracted when the generalization operation is selected, so that the description scope includes more non-sensitive information.

Advantages of the present invention:

The present invention utilizes the location information marked on the map, realizes the traversal and integration of the marked content in the geospatial data and the automatic processing of sensitive information desensitization, and overcomes the cumbersome work, low efficiency, error-prone and loopholes under the existing manual processing. The phenomenon.

Description of drawings

Fig. 1 is the working flow chart of the present invention.

Detailed ways

As shown in Figure 1, a KDTree-based image database data processing method includes the following steps

(1) Traverse and integrate map annotation information based on KDTree;

(2) Sensitive information detection based on word similarity;

(3) Desensitization based on the sensitivity of feature words.

The details are as follows:

(1) A method of traversing and integrating map annotation information based on KDTree:

This method takes advantage of the correlation of map annotation information in the location distribution, considers the points formed by all words, and greedily thinks that within a certain range, another point closest to the point where the current word is located can meet the conditions for integrating the two into a word , when the conditions for integrating into a word cannot be met, other points are considered within this range. The specific process is as follows:

Step 1: Extract the location coordinates and character content of all words in the attribute table on the map, and build KDTree according to the two-dimensional coordinates of each word. KDTree is a data structure that divides the k-dimensional data space, and is often used for range search and nearest neighbor search.

Step 2: Since people tend to read from top to bottom and from left to right, before integrating the marked content, arrange all the words in descending order of the ordinate, so that the words accessed later can be sorted by The sequence is added to the currently integrated field to obtain a preliminary queue Q. Create a flag array vis[] that records whether each word has been processed, initialized to 0. Traverse the queue Q until the queue is empty:

Step 3: If the current point p has been processed, that is, vis[p]=1, then jump to the next point in the queue; if vis[p]=0, search the range from the current single-word point p in KDTree to the threshold [ 0, ε], get the neighbor node set of the current unigram point p, and search for a unigram point q that satisfies the integration condition with the current unigram point p in the order of the distance point p from the nearest to the farthest in the neighbor node set, if successful If found, replace the current single-character point with q, and set vis[q] to 1; Step 4: Repeat step 3 until there is no single-character point that can be integrated with the current single-character point in the set of adjacent nodes, then the integrated single-character point As a mark, and according to the arrangement order of the words in the preliminary queue Q, process the next unprocessed word;

Step 5: Repeat steps 3 to 4 until all words in the preliminary queue Q are processed, and each label on the map is obtained.

In addition, when integrating an unprocessed point into a processed field, it is also necessary to satisfy that all the words in the formed new field s are on the same straight line and the distance between every two adjacent words is similar, that is, the distance between the two adjacent words on the map must be satisfied. The location distribution satisfies the conditions for forming a label, otherwise no integration is performed. During integration, there is almost no change in the ordinate, and the field with a large change in the abscissa should be placed at the front of the label content when merging, that is, the order from left to right.

After the queue Q is empty, all words are processed, and the label set obtained after integration is recorded as S={s ₁ ,s ₂ ,…,s _n }.

(2) A sensitive information detection method based on word similarity:

For the label set S, in order to protect data security, sensitive information detection based on word similarity is required, which mainly includes the following four steps:

Step 1: Use Chinese word segmentation technology to segment the marked content:

First, for each tag content _si of tag set S, it is converted into multiple word vectors by using Chinese word segmentation technology and word vector construction technology. The present invention uses jieba word segmentation to divide the map annotation into multiple words, and obtains s _i ={a ₁ ,a ₂ ,...,am }, where a ₁ ... _am are _m characteristic words obtained after division.

Step 2: Convert feature words and sensitive words into word vectors using word vector construction technology:

，取值范围为[0,1]，相似度越接近1表示两个词相似程度越大。The present invention uses word2vec to convert the feature words into word vectors, and the converted word vector of each feature word a _j is denoted as A _j . Similarly, convert all sensitive words b _k in the sensitive thesaurus into word vectors, denoted as B _k . At this time, the similarity between the feature word and the sensitive word can be quantified as the similarity between the feature word vector and the sensitive word vector, that is, taking the cosine of the angle between the inner product spaces of the two vectors as the similarity

, the value range is [0, 1], the closer the similarity is to 1, the greater the similarity between the two words.

Step 3: Sensitivity calculation of feature words:

，L值越大，则该敏感词的敏感程度越高。遍历敏感词库，对特征词a_j，定义其最大敏感度为In the sensitive thesaurus, each sensitive word _ck corresponds to a sensitivity level

, the larger the L value, the higher the sensitivity of the sensitive word. Traverse the sensitive thesaurus, define the maximum sensitivity of the feature word a _j as

表示敏感词库，计算每个敏感词向量与特征词向量的相似度与敏感级别的乘积，取其中的最大值表示特征词

的最大敏感度。in,

Represents a sensitive thesaurus, calculates the product of the similarity between each sensitive word vector and the feature word vector and the sensitivity level, and takes the maximum value to represent the feature word

maximum sensitivity.

大于θ时，该特征词

才具有敏感性，否则不认为是敏感词，其敏感度记为0，即特征词

的敏感度为Set the threshold parameter θ, when

When greater than θ, the feature word

Only have sensitivity, otherwise it is not considered a sensitive word, and its sensitivity is recorded as 0, that is, a characteristic word

The sensitivity is

Step 4: Sensitivity calculation of annotation content:

Since a label can be composed of multiple feature words, measuring the sensitivity of the label needs to consider the sensitivity of all its feature words. In addition, the sensitivity of the feature words that are often distributed near the end of the annotation content has a greater impact on the sensitivity of the annotation. For example, "Hunan Provincial Armed Police Hospital", although "armed police" is a sensitive word, the hospital is open to the public and is not a sensitive label. For example, "a certain nuclear power plant", "a certain military base", etc., the nouns at the end play a decisive role in the sensitivity of the labeling. Therefore, considering the location distribution of the feature words, the sensitivity of the annotation is defined as:

包含m个特征词，j表示

是该标注的第j个特征词，

为其敏感度，

为1,2,…,m的累加和。得出特征词与标注内容的敏感度后，可将标注内容按敏感度划分为高、中、低、非敏感4个级别。In the formula, label

Contains m feature words, j represents

is the j-th feature word of the annotation,

for its sensitivity ,

is the cumulative sum of 1,2,…,m. After obtaining the sensitivity of the feature words and the labeled content, the labeled content can be divided into four levels of high, medium, low, and insensitive according to the sensitivity.

(3) A desensitization method based on the sensitivity of feature words.

After detecting the sensitive data in the marked content according to method (2), the sensitive data needs to be processed according to the sensitivity level.

Step 1: For some labeled content that is not sensitive to the data itself, it may be mistakenly identified as sensitive data by the algorithm because it may contain feature words similar to sensitive words. For this, a whitelist of geographic labeling information can be constructed, and the algorithm is manually discovered each time. In case of wrongly identified non-sensitive data, it will be added to the white list to continuously improve the fault tolerance rate of the model of the present invention.

Step 2: To ensure the availability of map data, it is not appropriate to delete all sensitive data. After the sensitive data identified by the algorithm is filtered through the whitelist, some desensitization methods can be used to reduce its sensitivity to a non-sensitive level to meet the requirements of public disclosure. For the marked content with high sensitivity level, since it is very easy to threaten the security of geographic information after it is leaked, measures are taken to delete it directly. For labels with medium and low sensitivity levels, extract the sensitive feature words in them, randomly select desensitization methods such as deletion, replacement, and generalization for processing, and recalculate the labeling sensitivity. If it still does not meet the disclosure requirements after a certain number of iterations, delete the annotation.

Among them, when replacing, select the non-sensitive word with the greatest similarity to the current sensitive feature word in the non-sensitive word database as the replacement, and the generalization operation will abstract the specific content of the label description, so that the description scope includes more non-sensitive information, such as "PLA Logistics Base" is converted into "Warehouse" after generalization and replacement operations.

Claims (6)

1. A KDTree-based image database data processing method is characterized by comprising the following steps:

step one, traversing and integrating map labeling information based on KDTree to obtain a labeling set S = { S } formed by n labels₁,s₂,…,s_n}; sn represents the nth label;

the method specifically comprises the following steps:

1.1, extracting position coordinates and character contents of all single characters on a map in an attribute table of the map to form a data set, and constructing a KDTree according to a two-dimensional coordinate of each character;

1.2, arranging all the single characters according to the sequence of the ordinate from big to small to obtain a primary queue Q; creating a mark array vis [ ]) for recording whether each single character in the queue Q is processed or not, initializing the mark array to 0, and traversing the queue Q until the queue Q is empty;

1.3, processing the single characters in sequence according to the arrangement sequence of the single characters in the primary queue Q;

if the current single character point p is not processed, namely vis [ p ] =0, executing 1.4 steps, and juxtaposing vis [ p ] as 1;

if the current single sub-point p is processed, namely vis [ p ] =1, jumping to the next point in the queue Q;

1.4 searching points which are within a threshold value of [0, epsilon ] from the range of the current single character point p in the constructed KDTree to obtain a neighbor node set of the current single character point p, wherein epsilon represents a parameter of an integration range, and 1.5-2 times of the word width corresponding to the previous single character point p is taken; searching a single character point q meeting the integration condition with the current single character point p in the neighbor node set according to the sequence of the distance points p from near to far, if the single character point q is found successfully, replacing the current single character point with q, and juxtaposing vis [ q ] as 1; the single character point q is the point corresponding to the single character q in the KDTree;

1.5 repeating the step 1.4 until no single character point which can be integrated with the current single character point exists in the neighbor node set, and taking the integrated single character point as a label;

1.6 when no single character point which can be integrated with the current single character point exists in the neighbor node set; processing the next unprocessed single character according to the arrangement sequence of the single characters in the preliminary queue Q;

1.7 repeating the steps 1.3-1.6 until the single characters in the preliminary queue Q are all processed; obtaining each label on the map

Secondly, sensitive information detection based on word similarity is carried out on the label set S, and sensitivity grading is carried out on the map label content;

and thirdly, performing corresponding desensitization treatment according to the sensitivity level of the map labeling content.

2. The KDTree-based image database data processing method according to claim 1, wherein in step 1.4, the integration condition is as follows:

the single character point q is not processed, namely vis [ q ] = 0;

in the first case, when the integrated field only contains one current single character point p, the current single character point p is integrated with a single character point q which is closest to the previous single character point p in the neighbor node set; when only a single character point p exists in the neighbor node set, the single character point p forms a label;

and secondly, when the integrated field contains two or more words, namely, when the field formed by a plurality of words is integrated with the single word point q, judging whether all the words in the new field s formed by the single word point q and the integrated field are in the same straight line and whether the range R of the array formed by the distance between every two adjacent words meets the following conditions:

wherein Len represents the number of the single characters contained in the new field s,

indicates the new field siIndividual character point and jth individual character pointjThe distance of (a) to (b),

and

respectively represent taking the maximum value anda minimum value; gamma is 0.2-0.5 times of the width of the word in the new field s;

if all of the 2 integration conditions are satisfied, the fields formed by the plurality of words are integrated with the single word point q, and if at least one of the integration conditions is not satisfied, the fields formed by the plurality of words are not integrated with the single word point q.

3. The KDTree-based image database data processing method according to claim 2, wherein before integration, interference of duplicate words is first eliminated, if word frames corresponding to p and q intersect and the contents of the words are the same, then p and q are duplicate words, and q is deleted from the attribute table to realize deduplication.

4. The KDTree-based image database data processing method according to claim 1, wherein in step 1.7, for horizontally distributed labels, the individual characters in the label are arranged in order from small to large on the abscissa, and in order from left to right.

5. The KDTree-based image database data processing method of claim 1, wherein the second step comprises the steps of:

2.1: performing word segmentation on the marked content by adopting a Chinese word segmentation technology:

aiming at each piece of labeled content si of the label set S, converting the labeled content si into a plurality of word vectors by adopting a Chinese word segmentation technology and a word vector construction technology; obtaining si = { a1, a2, …, am }, wherein a1 … am is m feature words obtained after division;

2.2: converting the characteristic words and the sensitive words into word vectors by adopting a word vector construction technology:

converting the characteristic words into word vectors by using word2vec, and recording the word vectors after the j-th characteristic word Aj is converted as Aj; similarly, all sensitive words Bk in the sensitive word stock are converted into word vectors which are recorded as Bk; the similarity degree of the feature words and the sensitive words is quantized into the similarity degree of the feature word vectors and the sensitive word vectors, namely, the cosine value of an included angle of inner product spaces of the two vectors is taken as the similarity degree, the value range is [0,1], and the closer the similarity degree is to 1, the greater the similarity degree of the two words is;

representing the similarity between Aj and Bk;

2.3 sensitivity calculation of feature words:

in the sensitive word bank, each sensitive word ck corresponds to one sensitive level

，

The larger the value is, the higher the sensitivity degree of the sensitive word is; traversing the sensitive word stock, and defining the maximum sensitivity of the characteristic words aj for the characteristic words aj

Is composed of

Wherein,

representing a sensitive word bank, calculating the product of the similarity of each sensitive word vector and the feature word vector and the sensitivity level, and taking the maximum value to represent the maximum sensitivity of the feature words;

setting a threshold parameter theta when

When the value is larger than theta, the characteristic word aj has sensitivity, otherwise, the characteristic word aj is not considered as a sensitive word, the sensitivity is marked as 0, namely, the sensitivity of the characteristic word is

2.4 sensitivity calculation for annotated content:

defining the ith annotation

Sensitivity of (2)

Comprises the following steps:

in the formula, notation

Containing m feature words, j represents a label

The jth feature word of (1);

for marking

Sensitivity of, i.e. labelling

The sum of the sensitivity accumulation of the contained m characteristic words; after the sensitivities of the feature words and the labeled contents are obtained, the labeled contents are divided into 4 levels of high sensitivity, medium sensitivity, low sensitivity and non-sensitivity according to the sensitivities.

6. The KDTree-based image database data processing method of claim 1, wherein the third step comprises the steps of:

3.1, constructing a white list of the geographic marking information, and adding the non-sensitive data into the white list every time when the non-sensitive data which is wrongly identified by the algorithm is manually found, so that the fault tolerance rate is improved;

3.2 after the sensitive data are screened by the white list, the marked content with high sensitive level is directly deleted; labeling the medium and low sensitivity levels, extracting sensitive characteristic words in the medium and low sensitivity levels, randomly selecting a desensitization means of deletion, replacement and generalization for processing, then recalculating the desensitization labeled sensitivity, and completely deleting the corresponding label if the desensitization labeled sensitivity does not meet the public requirement after iteration preset times; when the non-sensitive word with the maximum similarity to the current sensitive characteristic word is selected as the replacement in the selection replacement, the specific content of the label description is abstracted when the generalization operation is selected, so that the description range comprises more non-sensitive information.

Images