CN101211341A - Image intelligent mode recognition and searching method - Google Patents

Abstract

The invention puts forward an image intelligent mode identification search method. The method can establish an image sample training set database and combine with basic text search engine technology and basic image content inquiry technology, so that a network creeper can perform Internet image search and URL information resolution, so as to catch the image URL and relevant information into a local primary database; perform such pre-processes as preliminary filtration, decompression and image pre-classification and etc for the images; then, calculate color characteristics, grain characteristics and shape characteristics of the extraction images, so as to gain corresponding characteristic vector sets; combine with the image URL information before saving the images into the image basic database and establishing an index for the images; perform characteristic vector similarity calculation for images in the image basic databases and sample training sets, and then, save the classified images into an image classification database; accept key words or image description that are input by the user, create the index vector, perform similarity calculation with the image characteristic vectors in the image classification database, and then, return the index results to the user.

Description

Image intelligent mode recognition and searching method

Technical field

The invention belongs to the computer picture area of pattern recognition, the rest image that dynamically generates in search at internet search engine has particularly proposed the searching method that carries out association analysis in conjunction with annotation of images text feature and image self metric.

Background technology

Image model identification is meant the research of view data being carried out the automated programming system design of recognition mode.An important applied field of pattern-recognition is to detect the text of particular type or image and operating process is controlled.Pattern (being also referred to as object, case or sample) is that " physics " of object is described, normally signal, image or simple numerical tabular.Feature be from pattern obtain to useful tolerance, attribute or the primitive of classifying.Such as, in image recognition, one 256 * 256 gray-scale map can obtain 65536 image metric values (light intensity), choose usually a small amount of metric as eigenwert improving recognition efficiency, but recognition correct rate has also descended thereupon.The target of pattern recognition system is to find a kind of mapping relations between representation space and version space.The basic skills of implementation pattern identification comprises: data clusters, organize data into the significant and useful data of respectively organizing with the method for certain similarity measurement, solution is a data-driven, do not rely on any supervised learning or guidance, advantage is that nicety of grading is higher, shortcoming is to realize more complicated, and efficiency ratio is lower.Statistical classification belongs to the supervised learning type, distributes based on obtain proper vector of all categories with probability statistics model, so that image is classified.The distribution that obtains proper vector is based on a training sample set (such as expert system) that classification is known, thereby train to learn how to classify with the sample set of known class label, realize simpler, travelling speed is than very fast, shortcoming be the data training sample set be difficult to accurately determine and nicety of grading lower.Neural network is inspired by the physiological knowledge of human brain tissue and found, and by forming a series of identical unit (neuron) that interknit, mutual contact can be transmitted enhancing or inhibition signal between different neurons.Strengthening or suppressing is by adjusting the weight coefficient realization of mutual contact.Opposite with the statistical classification method, neural network is irrelevant with " model ", shows the performance of sorter under a kind of unsupervised learning, and having can be by adjusting the advantage that makes that output approaches arbitrary target in feature space.The deficiency that exists, the one, mathematical expression is very complicated and do not have the versatility of application, and the 2nd, can not get the information of any semanteme from neural network itself.

At present, internet hunt no longer has been confined to text object, needs to carry out the search of multimedia files such as image, video more and more.Each large search engine such as Google, Yahoo, Live Search, Baidu, AltaVista etc. also provide the function of picture searching, but mainly are to search for according to the super chain URL of picture and the key word mark of picture.Two stages have been experienced in the retrieval of internet epigraph: the phase one is the retrieval based on key word.Subordinate phase is the retrieval based on the content of image self.In image indexing system based on key word, need earlier all images to be carried out the key word mark, could use global search technology that image is searched for then.There is the problem of two aspects in this method: the one, and this method needs more artificial participation, and along with the increase of picture number, this method is difficult to realize; Second problem is that the quantity of information that image comprises is huge, and different people is also inequality for the understanding of same image, and this just causes the unified standard of the mark neither one of image, thereby the result of retrieval can not meet user's demand well.Content-based retrieval is different from the retrieval based on key word, and it does not need too much artificial participation, and utilizes the feature (as color, texture, shape etc.) of image self to retrieve, and has stronger objectivity.But because these features real semantic information of representative image not, the content-based retrieval result is often unsatisfactory, sometimes incoherent advertisement bar, icon, Background, button is also retrieved out.In addition, if the key word mark does not exist, the real meaning of image is with unknown so.

Summary of the invention

In order to overcome the deficiency of conventional images search technique, the present invention proposes image intelligent mode recognition and searching method, the super chain URL of image is analyzed with the identification of image self content model combine, and be image interpolation key word automatically.

Technical scheme: the sample training collection database of at first setting up image by the human expert, again in conjunction with the text based search engine technique with based on the inquiring technology of picture material, carry out the search of the Internet images by new Web Crawler, resolve URL information, image URL and relevant information are grabbed in the local raw data base; To image carry out size, form tentatively filters, size is less, the undesirable image-erasing of form carries out necessary decompression, image pre-service such as presort; Calculate color characteristic, textural characteristics, the shape facility that extracts image then, obtain corresponding color, texture, shape facility vector set, combining image URL information is saved in the image basis database together, and sets up index for it; Image in the image basis database and sample training collection carry out the proper vector similarity and calculate, and are saved in the image classification database after image is classified; Accept the key word or the iamge description of user input, produce index vector, carry out similarity with characteristics of image vector in the image classification database and calculate, indexed results is returned to the user; If the user is unsatisfied with Search Results, then sample training collection database and image classification database are adjusted automatically, draw image searching result once more.

Useful technique effect is: the advantage that combines unsupervised learning and supervised learning in the pattern-recognition, improved the degree of hitting of picture searching in the large scale network, reduce the terrible response time that goes out the picture result for retrieval, and possess certain intelligent mode discriminator function.

Embodiment

Be to save storage space, improve transfer efficiency, the image on the internet is general to adopt forms such as JPEG, GIF through overcompression, TIFF, does not generally adopt the BMP form.Most popular on the internet at present is jpeg format, and JPEG (JointPhotographic Experts Group) is that ISO, CCITT, IEC are first international digital image compression standard of continuous tone still image joint research and development.Defined the system of two kinds of different performances in the Joint Photographic Experts Group: ultimate system and expanding system.Ultimate system becomes some 8 * 8 square with image segmentation, and each square is carried out dct transform, the DC coefficients by using one dimension DPCM coding after the quantification, and the AC coefficient adopts Run-Length Coding through Z scanning back.Use the Huffman encoding method to reduce redundance to the output of DPCM and Run-Length Coding at last, demoder is only stored two Huffman tables.Expanding system can provide the hierarchical operations pattern, and the coding of each picture content will be through repeatedly scanning, thereby the image of the different levels that spatial resolution increases progressively is provided, the predicted picture of each layer coding output as to the last layer picture coding time.Under the narrower situation of the network bandwidth, can advance images with the best in quality layer quickly.GIF (Graphics Interchange Format) is mainly data stream and a kind of transformat of designing, can be used for the sequential delivery and the demonstration of many images, and the animation icon that often occurs on webpage is this characteristic of having utilized GIF.GIF is divided into five parts in order: file header piece, logical screen descriptor, selectable color table (palette), video data block (or specific purposes piece) and tail piece, each piece is by identification code in each byte or condition code sign.What GIF adopted is the LZW compression algorithm, and this algorithm is with the symbol sebolic addressing in the string table preservation data stream, and each character string all has the index value of a coding as it.

Image is linked to html page by dual mode: embedded (InIine) and outer chain type (Reference).

1. embedded: this mode embeds html page by using the IMG element with image (being generally icon or small-sized image).The IMG element uses the SRC attribute to come self URL of specify image, uses the ALT attribute to specify the literal that image is described.Its basic format is:＜IMG SRC=" URL " ALT=" image description "

2. outer chain type: this mode realizes by anchor point element (Anchor element) A.The A element uses the HREF attribute to come the URL of designated links image.Its basic format is:＜A HREF=" URL " HYPERLINK TEXT＜/A 〉

The specific implementation method is as follows:

(1) makes up the sample training collection.According to user preferences and network picture present situation, select representational JPEG picture or GIF picture and be kept in the sorter, such as being divided into: N classes such as nature, building, personage, animal, plant, set up data set at each classification.Comprising such as " personage " class data set (Human): K image that contains standard faces comprises that the profile diagram of the earth four big ethnic group prototypes (H1, H2, H3 and H4) and each prototype (are designated as H respectively from different perspectives and apart from the figure that takes _Ij, wherein i refers to the sequence number of typical module, j refers to the sequence number of angle), prototype image has after treatment added piecewise linear outline line.What use when determining profile is a kind of dynamic profile tracking technique, constantly adjusts the quantity of line segment.Consider general situation, the textual description of selective extraction classified image, color, texture, shape are as proper vector C=C (Description, Color, the Texture of each classification, Shape), set up the initial characteristics of image vector sample training collection of each classification.Be provided with and the threshold value Tij (1＜=i＜=N, 1＜=j＜=4) of the Euclidean distance of each classifying image features vector and the initial weight coefficient Wj of 4 characteristics of image vectors, calculate the proper vector index of sample training collection $\mathrm{INDEX}={\mathrm{\Σ}}_{j=1}^4\left(\mathrm{WjCj}\right)$

(2) image creeping and obtaining.At the one or more new Web Crawler crawler of search engine server running background (or claiming network robot robot, Web Spider spider), the html web page of the website of certain limit is creeped.The network size of creeping when needs can adopt general depth-first algorithm or breadth-first algorithm to creep during less than setting value; When network size is big, can adopt concurrent frog jump heuristic search algorithm (seeing another patent of invention) etc. relatively to be fit to the higher algorithm of efficient that large scale network creeps carries out multithreading and creeps, can be according to the characteristics of object search, create two or more threads and the running priority level of each thread is set, control the execution of each thread, make the operational efficiency of the process of creeping higher.If only carry out creeping of image, then new Web Crawler is provided with the url filtering condition, resolve chained file and the descriptor that contains such as the URL of suffix such as * .jpg, * .jpeg, * .gif according to the image links mode of html page.Adopt the timestamp technology,, then image file and the text annotated information that is linked in the page stored in the local raw data base if the picture material on the discovery webpage changes or new image webpage occurs.

(3) extraction of characteristics of image.In order to improve entire system identification travelling speed, consider to use the characteristics of image identical as four category feature vectors (textual description Description, color Color, texture Texture, shape Shape) with the sample training collection.Reading images from local raw data base connects into character string as " textual description " proper vector with text messages such as the note of image, URL addresses.Be the Description=annotation of images || the URL address.

The extraction of other category feature vectors is as follows:

1. " color " (Color) THE ACQUISITION OF FEATURE VECTOR.Color characteristic space commonly used comprises: the one, and the most basic RGB feature space, based on Descartes's three-dimensional coordinate system, be a cube shaped, three axles are represented R (red redness), G (green green) and B (blue blueness) respectively, and each span all is 0 to 255.True origin (0,0,0) expression " black ", (255,0,0) expression " redness ", (0,255,0) expression " green ", (0,0,255) expressions " blueness ", (255,255,255) expressions " white ", from " black " to " white " continuous cubical diagonal line is represented the Continuous Gray Scale value from " black " to " white ", and other each points are represented different colors in the cube.According to this spatial model, every width of cloth coloured image can be broken down into 3 independently on the plane.The 2nd, through the feature space of linear transformation, comprise the YUV color space that is used for the pal mode color standard, be used for the YIQ color space of TSC-system formula color video standard and be used for the YCrCb color space of JPEG compression graphical format.The 3rd, through CIE (LUV) color space of linear transformation and nonlinear transformation.The 4th, the hsv color space, HSV represents three kinds of attributes of color respectively: tone (Hue), saturation degree (Saturation) and brightness (Value).Traditional RGB method for expressing is realized very simple, but not too satisfies the visual characteristic of human eye and the specific (special) requirements of picture on the Internet.The HSV method for expressing satisfies the human eye characteristic most, but realization is very complicated, need RSB be converted to HSV by a large amount of computings.Consider the problem of the extensive image recognition in the Internet, and picture format mostly is the YCrCb method for expressing greatly.So propose faster implementation trade-off way of a kind of ratio, at first judge the form of image, if be JPEG (JPG) form, " color " proper vector is expressed as the one dimension transition form of tri-vector Y, Cr, Cb, i.e. color feature vector $\mathrm{Color}=\frac{Y+\mathrm{Cr}+\mathrm{Cb}}{3},$ Y=0.299R+0.587G+0.114B wherein, Cr=0.5R-0.4187G-0.0813B, Cb=-0.1687R-0.3313G+0.5B; If be the BMP form, then $\mathrm{Color}=\frac{R+G+B}{3}.$ 2. " texture " (Texture) THE ACQUISITION OF FEATURE VECTOR.Textural characteristics is the basic structure of expression vision, mainly comprise roughening, directivity, contrast, periodically, concavity and convexity etc.Typical textural characteristics comprises the Tamura textural characteristics, based on the textural characteristics of wavelet transformation, co-occurrence autoregression textural characteristics etc.The Texture Segmentation of image is a suitable difficulty and the very big task of calculated amount.Therefore, the present invention proposes fairly simple implementation method, only calculates roughness, contrast and directivity.The calculating of roughness: 1) calculate moving average (moving average), for 2 ^k* 2 ^kWindow, moving average is:

$a_k(i,j)=\underset{i^{\′}=i-2^{k-1}}{\overset{i+2^{k-1}-1}{\mathrm{\Σ}}}\underset{j^{\′}=j-2^{k-1}}{\overset{j+2^{k-1}-1}{\mathrm{\Σ}}}\frac{p(i^{\′},j^{\′})}{2^{2k}}$

2) calculated level with vertical to deviation

c _k(i，j)＝max(|a _k(i-2 ^k-1，j)-a _k(i+2 ^k-1，j)|，|a _k(i，j-2 ^k-1)-a _k(i，j+2 ^k-1)|)

3) determine window size

$\hat{k}(i,j)=\arg \underset{k}{\max c_k}(i,j)$

4) calculate average window

The calculating of contrast:

$\mathrm{contrast}=\frac{\mathrm{\σ}}{\sqrt[4]{{\mathrm{\α}}_4}}$

Directivity is meant the direction of gray-scale value in the image.Following four steps of calculated direction sexual needs:

1) calculates the gradient of each pixel.Gradient refers to this pixel the fastest direction of gray-scale value increase on every side.Horizontal gradient equals the deviation between three gray-scale values of three gray-scale values of leftmost pixel and right pixels, and VG (vertical gradient) then is three gray-scale value deviations of pixel up and down.

2) polar coordinates of compute gradient vector

$\left(\right|g|,\mathrm{\φ})=(\frac{\left|{\mathrm{\Δ}}_h\right|+\left|{\mathrm{\Δ}}_v\right|}{2},{\tan }^{-1}\left(\frac{{\mathrm{\Δ}}_v}{{\mathrm{\Δ}}_h}\right)+\frac{\mathrm{\π}}{2})$

3) histogram of calculating tilt vector angle

n _φ(k): expression is satisfied $\frac{2k-1}{2n}<\frac{\mathrm{\&phi;}}{\mathrm{\&pi;}}<\frac{2k+1}{2n}\left(\mathrm{mod}1\right)$ With | the ratio of the pixel of g|＞t condition.

4) obtain after the histogram, calculate the variation summation of crest (trough is to trough) value on every side

Directivity (Direction)=trough is to the summation that changes between the trough

Consider the large percentage that roughness and contrast account for the human eye sense organ in the textural characteristics, so texture feature vector

$\mathrm{Texture}=\frac{2\mathrm{Coarseness}+2\mathrm{Contrast}+\mathrm{Direction}}{5}$

3. " shape " (Shape) THE ACQUISITION OF FEATURE VECTOR.It mainly is extraction at the picture edge characteristic of distinguishing easily.Edge detection operator is checked the neighborhood of each pixel, and rate of gray level is quantized, and also comprises determining of direction.Great majority use the method for asking convolution based on the directional derivative mask, and differentiating in the practical application is to utilize the difference approximation differential to carry out.Several edge detection methods commonly used comprise Sobel edge detection operator method, Prewitt edge detection operator method, Gauss-Laplace operator detection method, Canny edge detection method.The method that the present invention adopts length, width, rectangle degree, the circularity of fairly simple practicality to combine.Calculate the minimum and maximum coordinate figure of object boundary point, just can obtain level (L) and the vertical span (W) and the ratio r=L/W between them of object, such boundary rectangle is the boundary rectangle (MER-MinimumEnclosing Rectangle) of object minimum.The rectangle degree reflects the full level of object to its boundary rectangle, i.e. R=A/Amer with the portrayal recently of the area of the area boundary rectangle minimum with it of object.Circularity is used for portraying the complexity of object boundary, and they get minimum value when circular boundary.The most frequently used circularity be girth square with the ratio of area, i.e. C=P ²/ A." shape " proper vector formula is as follows:

Shape＝C/r+r/R+R/C＝(RC ²+Cr ²+rR ²)/rRC

At last, obtain image proper vector E=E (Description, Color, Texture, Shape).The image that will have proper vector is set up index $\mathrm{index}={\mathrm{\&Sigma;}}_{j=1}^4\left(\mathrm{WjEj}\right)$ After be saved in the image basis database.

(4) classification of image.Reading images from the image basis database is obtained the Euclidean distance that the proper vector of image and sample training are concentrated each characteristic of division vector, i.e. Dij=|Eij-Cij|, 1＜=i＜=N wherein, 1＜=j＜=4.Similarity function is $\mathrm{Fi}=\left({\mathrm{\&Sigma;}}_{j=1}^4\frac{\mathrm{WjDij}}{\mathrm{Tij}}\right)/\left({\mathrm{\&Sigma;}}_{j=1}^4\mathrm{Wj}\right),$ 1＜＝i＜＝N，1＜＝j＜＝4。Wherein, Wj is the weight coefficient of characteristics of image vector, can adjust automatically according to user's input.After the similarity with each classification samples training set of image is obtained, obtain minimum value Fk=min (Fi) again.With image through hash functional transformation h=(index) mod (INDEX (k)) after, with image, characteristics of image vector and k class training sample proper vector copy to corresponding image classification database in, the sequence number of this classification is h.To the image of storage according to ascending ordering of similarity function value F (value is more little, shows that image is similar more to training sample, and it is also forward more to sort) of training sample proper vector, sequence number is p, the index that image is new is h+p.And, generate the corresponding thumbnail (user distinguishes for convenience) of image according to dimension scale.

(5) user search.Representing to the user at the web client end interface is the picture search compound condition, comprise that pictograph describes the respective weights coefficient (Wj) of S1, color S2, texture S3 and shape S4 and condition ' etc., selection according to the user, submit to compound condition to calculate for the web server, obtain user's search key $\mathrm{key}={\mathrm{\&Sigma;}}_{j=1}^4\left({\left(\mathrm{Wj}\right)}^{,}\mathrm{Sj}\right).$ Then, submit to the database search program on backstage, at first with sample training collection database in the proper vector index carry out Euclidean distance and calculate, find out the minimum value min of distance (| key-INDEX (m) |), the classification number of this minimum value correspondence is m.Then, behind key process hash functional transformation 1=(key) mod (INDEX (m)), find the 1st class image in the image classification database, the image searching result that hits is returned to the user according to similarity degree, return results information comprises image feature information and thumbnail.If user's Search Results is dissatisfied, then can changes the search condition weight coefficient and submit searching request once more to.Search engine backstage image processing program then calculates sample training collection proper vector index again according to the weight coefficient that the user submits to, reclassify according to (4) step after characteristics of image vector index in the image classification device recomputates then, at last the result is returned to the user, till the user no longer submits to weight coefficient to revise.

Claims (2)

1. image intelligent mode recognition and searching method, it is characterized in that, at first set up the sample training collection database of image by the human expert, again in conjunction with the text based search engine technique with based on the inquiring technology of picture material, carry out the search of the Internet images by new Web Crawler, resolve URL information, image URL and relevant information are grabbed in the local raw data base; To image carry out size, form tentatively filters, size is less, the undesirable image-erasing of form carries out necessary decompression, image pre-service such as presort; Calculate color characteristic, textural characteristics, the shape facility that extracts image then, obtain corresponding color, texture, shape facility vector set, combining image URL information is saved in the image basis database together, and sets up index for it; Image in the image basis database and sample training collection carry out the proper vector similarity and calculate, and are saved in the image classification database after image is classified; Accept the key word or the iamge description of user input, produce index vector, carry out similarity with characteristics of image vector in the image classification database and calculate, indexed results is returned to the user; If the user is unsatisfied with Search Results, then sample training collection database and image classification database are adjusted automatically, draw image searching result once more.

2. according to the described a kind of image intelligent mode recognition and searching method of claim 1, it is characterized in that,

(1) makes up the sample training collection; According to user preferences and network picture present situation, select representational JPEG picture or GIF picture and be kept in the sorter, such as being divided into: N classes such as nature, building, personage, animal, plant, set up data set at each classification; Comprising such as " personage " class data set (Human): K image that contains standard faces comprises that the profile diagram of the earth four big ethnic group prototypes and each prototype are designated as H respectively from different perspectives and apart from the figure that takes _Ij, wherein i refers to the sequence number of typical module, and j refers to the sequence number of angle, and prototype image has after treatment added piecewise linear outline line; What use when determining profile is a kind of dynamic profile tracking technique, constantly adjusts the quantity of line segment; Consider general situation, the textual description of selective extraction classified image, color, texture, shape are set up the initial characteristics of image vector sample training collection of each classification as the proper vector C=C of each classification; Be provided with and the threshold value Tij (1＜=i＜=N, 1＜=j＜=4) of the Euclidean distance of each classifying image features vector and the initial weight coefficient Wj of 4 characteristics of image vectors, calculate the proper vector index of sample training collection $\mathrm{INDEX}={\mathrm{\&Sigma;}}_{j=1}^4\left(\mathrm{WjCj}\right)$

(2) image creeping and obtaining; At the one or more new Web Crawler crawler of search engine server running background (or claiming network robot robot, Web Spider spider), the html web page of the website of certain limit is creeped; The network size of creeping when needs can adopt general depth-first algorithm or breadth-first algorithm to creep during less than setting value; When network size is big, can adopt concurrent frog jump heuristic search algorithm etc. relatively to be fit to the higher algorithm of efficient that large scale network creeps carries out multithreading and creeps, can be according to the characteristics of object search, create two or more threads and the running priority level of each thread is set, control the execution of each thread, make the operational efficiency of the process of creeping higher; If only carry out creeping of image, then new Web Crawler is provided with the url filtering condition, resolve chained file and the descriptor that contains such as the URL of suffix such as * .jpg, * .jpeg, * .gif according to the image links mode of html page; Adopt the timestamp technology,, then image file and the text annotated information that is linked in the page stored in the local raw data base if the picture material on the discovery webpage changes or new image webpage occurs;

(3) extraction of characteristics of image; In order to improve entire system identification travelling speed, consider to use the characteristics of image identical as four category feature vectors (textual description Description, color Color, texture Texture, shape Shape) with the sample training collection; Reading images from local raw data base connects into character string as " textual description " proper vector with text messages such as the note of image, URL addresses; Be the Description=annotation of images || the URL address;

The extraction of other category feature vectors is as follows:

1. " color " (Color) THE ACQUISITION OF FEATURE VECTOR; Color characteristic space commonly used comprises: the one, and the most basic RGB feature space, based on Descartes's three-dimensional coordinate system, be a cube shaped, three axles are represented R (red redness), G (green green) and B (blue blueness) respectively, and each span all is 0 to 255; True origin (0,0,0) expression " black ", (255,0,0) expression " redness ", (0,255,0) expression " green ", (0,0,255) expressions " blueness ", (255,255,255) expressions " white ", from " black " to " white " continuous cubical diagonal line is represented the Continuous Gray Scale value from " black " to " white ", and other each points are represented different colors in the cube; According to this spatial model, every width of cloth coloured image can be broken down into 3 independently on the plane; The 2nd, through the feature space of linear transformation, comprise the YUV color space that is used for the pal mode color standard, be used for the YIQ color space of TSC-system formula color video standard and be used for the YCrCb color space of JPEG compression graphical format; The 3rd, through CIE (LUV) color space of linear transformation and nonlinear transformation; The 4th, the hsv color space, HSV represents three kinds of attributes of color respectively: tone (Hue), saturation degree (Saturation) and brightness (Value); Traditional RGB method for expressing is realized very simple, but not too satisfies the visual characteristic of human eye and the specific (special) requirements of picture on the Internet; The HSV method for expressing satisfies the human eye characteristic most, but realization is very complicated, need RSB be converted to HSV by a large amount of computings; Consider the problem of the extensive image recognition in the Internet, and picture format mostly is the YCrCb method for expressing greatly; So propose faster implementation trade-off way of a kind of ratio, at first judge the form of image, if be JPEG (JPG) form, " color " proper vector is expressed as the one dimension transition form of tri-vector Y, Cr, Cb, i.e. color feature vector $\mathrm{Color}=\frac{Y+\mathrm{Cr}+\mathrm{Cb}}{3},$ Y=0.299R+0.587G+0.114B wherein, Cr=0.5R-0.4187G-0.0813B, Cb=-0.1687R-0.3313G+0.5B; If be the BMP form, then $\mathrm{Color}=\frac{R+G+B}{3};$ 2. " texture " (Texture) THE ACQUISITION OF FEATURE VECTOR; Textural characteristics is the basic structure of expression vision, mainly comprise roughening, directivity, contrast, periodically, concavity and convexity etc.; Typical textural characteristics comprises the Tamura textural characteristics, based on the textural characteristics of wavelet transformation, co-occurrence autoregression textural characteristics etc.; The Texture Segmentation of image is a suitable difficulty and the very big task of calculated amount; Therefore, the present invention proposes fairly simple implementation method, only calculates roughness, contrast and directivity; The calculating of roughness: 1) calculate moving average (moving average), for 2 ^k* 2 ^kWindow, moving average is:

$a_k(i,j)=\underset{i^{\&prime;}=i-2^{k-1}}{\overset{i+2^{k-1}-1}{\mathrm{\&Sigma;}}}\underset{j^{\&prime;}=j-2^{k-1}}{\overset{j+2^{k-1}-1}{\mathrm{\&Sigma;}}}\frac{p(i^{\&prime;},j^{\&prime;})}{2^{2k}}$

2) calculated level with vertical to deviation

c _k(i，j)＝max(|a _k(i-2 ^k-1，j)-a _k(i+2 ^k-1，j)|，|a _k(i，j-2 ^k-1)-a _k(i，j+2 ^k-1)|)

3) determine window size

$\hat{k}(i,j)=\arg \underset{k}{\max c_k}(i,j)$

4) calculate average window

The calculating of contrast:

$\mathrm{contrast}=\frac{\mathrm{\&sigma;}}{\sqrt[4]{{\mathrm{\&alpha;}}_4}}$

Directivity is meant the direction of gray-scale value in the image; Following four steps of calculated direction sexual needs:

1) calculates the gradient of each pixel; Gradient refers to this pixel the fastest direction of gray-scale value increase on every side; Horizontal gradient equals the deviation between three gray-scale values of three gray-scale values of leftmost pixel and right pixels, and VG (vertical gradient) then is three gray-scale value deviations of pixel up and down;

Gradient calculation $g=\left(\begin{array}{c}{\mathrm{\&Delta;}}_h\\ {\mathrm{\&Delta;}}_v\end{array}\right)$

Horizontal gradient ${\mathrm{\&Delta;}}_h=\underset{k\&Element;\{-\mathrm{1,0,1}\}}{\mathrm{\&Sigma;}}p(i+1,j+k)-p(i-1,j+k)$

VG (vertical gradient) ${\mathrm{\&Delta;}}_v=\underset{k\&Element;\{-\mathrm{1,0,1}\}}{\mathrm{\&Sigma;}}p(i+k,j+1)-p(i+k,j-1)$

2) polar coordinates of compute gradient vector

$\left(\right|g|,\mathrm{\&phi;})=(\frac{\left|{\mathrm{\&Delta;}}_h\right|+\left|{\mathrm{\&Delta;}}_v\right|}{2},{\tan }^{-1}\left(\frac{{\mathrm{\&Delta;}}_v}{{\mathrm{\&Delta;}}_h}\right)+\frac{\mathrm{\&pi;}}{2})$

3) histogram of calculating tilt vector angle

n _φ(k): expression is satisfied $\frac{2k-1}{2n}<\frac{\mathrm{\&phi;}}{\mathrm{\&pi;}}<\frac{2k+1}{2n}\left(\mathrm{mod}1\right)$ With | the ratio of the pixel of g|＞t condition;

4) obtain after the histogram, calculate the variation summation of crest (trough is to trough) value on every side

Directivity (Direction)=trough is to the summation that changes between the trough

Consider the large percentage that roughness and contrast account for the human eye sense organ in the textural characteristics, so texture feature vector

$\mathrm{Texture}=\frac{2\mathrm{Coarseness}+2\mathrm{Contrast}+\mathrm{Direction}}{5}$

3. " shape " (Shape) THE ACQUISITION OF FEATURE VECTOR; It mainly is extraction at the picture edge characteristic of distinguishing easily; Edge detection operator is checked the neighborhood of each pixel, and rate of gray level is quantized, and also comprises determining of direction; Great majority use the method for asking convolution based on the directional derivative mask, and differentiating in the practical application is to utilize the difference approximation differential to carry out; Several edge detection methods commonly used comprise Sobel edge detection operator method, Prewitt edge detection operator method, Gauss-Laplace operator detection method, Canny edge detection method; The method that the present invention adopts length, width, rectangle degree, the circularity of fairly simple practicality to combine; Calculate the minimum and maximum coordinate figure of object boundary point, just can obtain level (L) and the vertical span (W) and the ratio r=L/W between them of object, such boundary rectangle is the boundary rectangle (MER-MinimumEnclosing Rectangle) of object minimum; The rectangle degree reflects the full level of object to its boundary rectangle, i.e. R=A/Amer with the portrayal recently of the area of the area boundary rectangle minimum with it of object; Circularity is used for portraying the complexity of object boundary, and they get minimum value when circular boundary; The most frequently used circularity be girth square with the ratio of area, i.e. C=P ²/ A; " shape " proper vector formula is as follows:

Shape＝C/r+r/R+R/C＝(RC ²+Cr ²+rR ²)/rRC

At last, obtain image proper vector E=E (Description, Color, Texture, Shape); The image that will have proper vector is set up index $\mathrm{index}={\mathrm{\&Sigma;}}_{j=1}^4\left(\mathrm{WjEj}\right)$ After be saved in the image basis database;

(4) classification of image; Reading images from the image basis database is obtained the Euclidean distance that the proper vector of image and sample training are concentrated each characteristic of division vector, i.e. Dij=|Eij-Cij|, 1＜=i＜=N wherein, 1＜=j＜=4; Similarity function is $\mathrm{Fi}=\left({\mathrm{\&Sigma;}}_{j=1}^4\frac{\mathrm{WjDij}}{\mathrm{Tij}}\right)/\left({\mathrm{\&Sigma;}}_{j=1}^4\mathrm{Wj}\right),$ 1＜=i＜=N, 1＜=j＜=4; Wherein, Wj is the weight coefficient of characteristics of image vector, can adjust automatically according to user's input; After the similarity with each classification samples training set of image is obtained, obtain minimum value Fk=min (Fi) again; With image through hash functional transformation h=(index) mod (INDEX (k)) after, with image, characteristics of image vector and k class training sample proper vector copy to corresponding image classification database in, the sequence number of this classification is h; To the image of storage according to ascending ordering of similarity function value F of training sample proper vector, sequence number is p, the index that image is new is h+p; And, generate the corresponding thumbnail of image according to dimension scale;

(5) user search; Representing to the user at the web client end interface is the picture search compound condition, comprise that pictograph describes the respective weights coefficient (Wj) of S1, color S2, texture S3 and shape S4 and condition ' etc., selection according to the user, submit to compound condition to calculate for the web server, obtain user's search key $\mathrm{key}={\mathrm{\&Sigma;}}_{j=1}^4\left({\left(\mathrm{Wj}\right)}^{,}\mathrm{Sj}\right);$ Then, submit to the database search program on backstage, at first with sample training collection database in the proper vector index carry out Euclidean distance and calculate, find out the minimum value min of distance (| key-INDEX (m) |), the classification number of this minimum value correspondence is m; Then, behind key process hash functional transformation l=(key) mod (INDEX (m)), find l class image in the image classification database, the image searching result that hits is returned to the user according to similarity degree, return results information comprises image feature information and thumbnail; If user's Search Results is dissatisfied, then can changes the search condition weight coefficient and submit searching request once more to; Search engine backstage image processing program then calculates sample training collection proper vector index again according to the weight coefficient that the user submits to, reclassify according to (4) step after characteristics of image vector index in the image classification device recomputates then, at last the result is returned to the user, till the user no longer submits to weight coefficient to revise.