CN103336835B - Image retrieval method based on weight color-sift characteristic dictionary - Google Patents

Abstract

An image retrieval method based on a weight color-sift feature dictionary, which includes randomly selecting a training image in an image to be retrieved, extracting the edge of the training image, and extracting the color-sift feature of all training image edge points, and Build a feature dictionary with the color-sift feature; input the image to be retrieved and extract the color-sift feature of the edge point of the search image and the image to be retrieved, and extract the weight histogram feature of the search image and the image to be retrieved based on the feature dictionary ; Perform similarity matching based on the weight histogram feature between the retrieved image and the image to be retrieved in the database; check whether to traverse all the images to be retrieved in all databases, if so, display the image retrieval result according to the similarity matching result, if not , re-do similarity matching. Improved accuracy and callback rates when applied to retrieval of large image databases. The algorithm is invariant to scale, translation and rotation. It has the characteristics of locality, particularity, quantity and efficiency.

Description

Image retrieval method based on weighted color-sift feature dictionary

technical field

The invention belongs to the technical field of image retrieval, and specifically relates to an image retrieval method based on a weight color-sift feature dictionary, which is based on image content and image feature extraction to realize the process of image analysis and retrieval, and allows users to input a or multiple images to find other images with the same or similar content.

Background technique

An image is a similar and vivid description or portrait of an objective object. In other words, an image is a representation of an objective object, which contains information about the object being described. It is people's primary source of information. According to statistics, about 75% of the information a person obtains comes from vision. As the saying goes, "seeing is better than hearing a hundred times" and "clear at a glance", both reflect the unique effect of images in information transmission. How to quickly and accurately extract image content is the most critical step in image retrieval. SIFT is a feature detection method that summarizes invariant technology in 2004 by David G Lowe, and proposes an image local feature description operator that is invariant to scale space, image scaling, rotation, and even affine. The so-called image content-based retrieval refers to searching for images containing specific objects from the image library, and also includes retrieving video clips containing specific objects from continuous video images. It is different from the traditional image retrieval method. The present invention proposes an image retrieval method based on the weight color-sift feature dictionary, which integrates the weight color-sift feature dictionary technology, thereby providing a more effective retrieval method. Can be applied to digital libraries, medical diagnosis, image classification, WEB related applications, public safety and criminal investigation, etc.

The patent application "An Image Retrieval Method Based on Sketch Feature Extraction" (Patent Application No. 201110196051.2, Publication No. 201110196051.2) filed by Tsinghua University discloses an image retrieval method based on sketch feature extraction, which involves the field of image retrieval. The method comprises the steps of: extracting a training feature vector to obtain a feature dictionary; extracting an input feature vector to obtain an input feature vector set, performing a counting operation on the feature dictionary to obtain an input feature frequency vector, and then obtaining an interesting feature word and a non-interest feature word; Extract the retrieval feature vector, obtain the retrieval feature vector set, and then obtain the retrieval feature frequency vector; then obtain the interesting retrieval feature frequency vector, the non-interest retrieval feature frequency vector, the interesting input feature frequency vector and the non-interest input feature frequency vector; and then calculate the input sketch The similarity with each retrieval sketch is outputted as a retrieval result. This method has good user interaction and improves the efficiency and accuracy of image retrieval, but the feature dimension is large, and two features of interest and non-interest are considered, resulting in low retrieval efficiency and slow speed when applied to large databases .

The patent application "An interactive image retrieval method combining user evaluation and annotation" (patent application number 201310128036.3, publication number CN103164539A) filed by Communication University of China discloses an interactive image retrieval method combining user evaluation and annotation, which belongs to multimedia the field of information retrieval. This method uses a comprehensive retrieval method based on the combination of physical characteristics of images and text. During the retrieval process, users are allowed to describe the query image in text, or select keywords provided by the system. For the relevant evaluation of "unsatisfied", the image retrieval system automatically marks the relevant satisfactory images marked by the user to form high-level semantic information; with the continuous use of the user, the system will generate a rich semantic information database. Considering the differences in text annotations of the same picture by different users and by the same user at different times, the present invention combines the user's credibility in the process of generating the semantic information database. When searching, the query image with semantic information is retrieved using a comprehensive retrieval method based on the combination of features and text, which improves the accuracy of the retrieval results. Although this method combines the interactive image retrieval method of user evaluation and annotation to obtain high-level semantic information and improve the accuracy of real-time retrieval, when it is applied to a large image database, due to the complicated similarity metrics of various images, the processing is complicated. The labeling increases the computational complexity and reduces the retrieval efficiency of the image, resulting in a low callback rate and retrieval rate of the returned retrieval result set.

Contents of the invention

The present invention aims at the deficiencies of the above-mentioned prior art, and proposes an image retrieval method based on a weighted color-sift feature dictionary, which improves retrieval efficiency, speed and callback rate when large databases are used.

An image retrieval method based on weight color-sift feature dictionary, which includes,

Randomly select a training image in the image to be retrieved, extract the edge of the training image, and extract the color-sift feature of all training image edge points, and construct a feature dictionary with the color-sift feature;

Input the image to be retrieved and extract the color-sift feature of the image to be retrieved and the edge point of the image to be retrieved, based on the feature dictionary to extract the weight histogram feature of the image to be retrieved and the image to be retrieved; Similarity matching based on weight histogram features;

Detect whether to traverse all the images to be retrieved in all databases, if yes, display the image retrieval results according to the similarity matching results, if not, perform similarity matching again.

On the basis of the above technical solution, randomly selecting training images from the images to be retrieved includes randomly selecting one image for each type of image in the image database to be retrieved to form the training image database.

On the basis of the above technical solution, extracting the edge of the training image includes: extracting the edge of the training image includes randomly selecting a training image in the training image database for gray scale transformation, and processing it through a direction adjustable filter , select the two-dimensional Gaussian function as the filter kernel function, get the energy function W _σ (x,y,θ) of each pixel point in 2L directions, and extract the significant pixel points of the edge of the image through threshold judgment, where L represents the direction The number of , x and y represent the coordinate value of the pixel point, θ is the value of the direction, the range is 0~2π, and the interval is π/L.

On the basis of the above technical solution, extracting the color-sift feature of all training image edge points includes: according to the edge salient pixels of the selected training image, the original training color image is respectively divided into red-R, green-G and blue-B The color-sift features are extracted by three channels, and the color-sift features dec _r (e), dec _g (e) of the red-R, green-G and blue-B three channels of the red-R, green-G and blue-B channels of each prominent pixel point of the image are obtained and dec _b (e), e represents the e-th edge point of the image, e=1, 2,..., E, E is the total number of all edge salient pixels of the image.

On the basis of the above technical solution, the extraction of the color-sift feature of all training image edge points includes: extracting the color-sift feature of all edge salient pixels of each training image in the training image database, traversing the training image database All images of , whose features are sequentially dec _r,m (e _m ), dec _g,m (e _m ) and dec _b,m (e _m ) in the three color channels, m=1,2,... , M, M is the size of the training image database, em represents the em edge pixel of the _{mth training image, em = 1,2,...,E m , E m} is all the _mth training images The total number of edge salient pixels.

On the basis of the above-mentioned technical scheme, the step of constructing a feature dictionary includes the color-sift feature dec _r,m (e _m ), dec _g,m (e _m ) and dec _b,m (e _m ), through K-means clustering calculation, take K cluster centers to get red-R channel, green-G channel and blue-B channel The two-dimensional feature dictionaries cod _r , cod _g and cod _b of w rows and K columns, where K is the number of K-means cluster centers, that is, the size of the dictionary.

On the basis of the above technical solution, the weight histogram feature X includes the weight histogram features of the three channels of red-R, green-G and blue-B, and the retrieval image and the image to be retrieved are calculated based on the weight Similarity matching of histogram features, calculate 2-norm similarity distance to get D _i (X,X′ _i ).

On the basis of the above-mentioned technical solution, the weight histogram feature based on the color-sift feature dictionary of the retrieved image is obtained by encoding and calculating the retrieved image, including the following steps:

1) Retrieve the color-sift feature sec _r (o) of all edge significant pixels on the red channel R of the image, and calculate sec _r (o) corresponding to the Euclidean distance of the K cluster centers of the red channel two-dimensional feature dictionary cod _r , select the clustering center with the smallest distance value as the center of the edge point, and perform first-order distance statistical calculation on all edge significant points to obtain the frequency histogram histogram corresponding to the two-dimensional feature dictionary cod _r of the retrieved image in the red channel _;

2) Assuming that q _k edge significant points fall in the _k -th cluster center, calculate the centrifugal weight li( k), calculate the maximum centrifugal weight of all significant points in the cluster center to obtain the weight vector α(k) of the cluster center, and make a matrix for the frequency histogram his _r of the retrieved image and the corresponding weight vector α(k) The dot multiplication operation obtains the weight vector hst _r of the retrieved image, that is, the multiplication of corresponding elements, hst _r is a K-dimensional column vector, k represents the k-th cluster center, and the value is 1, 2,...,K, K represents the number of cluster centers, that is, the size of the dictionary;

3) Do the same calculation as the red channel R in the green channel G and blue channel B, and finally get the weight vector hst _g of the image green channel G and the weight vector hst _b of the blue channel B, for the three channels The weight vector is integrated and calculated to obtain the weight histogram feature X based on the color-sift feature dictionary of the retrieved image.

On the basis of the above technical solution, the q _k edge salient points clustered to the kth cluster center are calculated to calculate the centrifugal weight li(k) of the salient point for the cluster center, and the cluster The weight vector α(k) of the cluster center is obtained by the largest centrifugal weight of all the significant points in the center, which is calculated by the following formula:

$\mathrm{<span\; class="notranslate">\; li</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}}\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; cod</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; cod</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; ,</span>$

α(k)=max(li(u,k)),

Among them, u represents the u-th edge significant point falling in the k-th cluster center, and the value is 1, 2,...,q _k , k represents the k-th cluster center, and the value is 1, 2,. .., K, K represents the number of cluster centers, that is, the size of the dictionary, and q _k represents the total number of edge significant points falling on the kth cluster center.

Compared with the prior art, the present invention uses a direction-adjustable filter for edge extraction, which can effectively determine the direction of the main direction of each pixel edge, and then quickly and accurately extract the edge pixel information of the image through threshold determination , the image edge pixel information obtained by extraction can quickly and accurately perform feature extraction in the next step, which improves the speed and accuracy of retrieval when applied to real-time human-computer interaction and large image databases. A retrieval strategy combining the color-sift feature of pixel points in the edge direction and the coding dictionary is adopted to extract the color-sift feature of the significant pixel points on the edge and calculate the weight histogram feature through the coding dictionary constructed based on the weight color-sift feature. The representation of images is more typical, and can more effectively represent the feature differences of images. When applied to the retrieval process, it improves the accuracy and callback rate when applied to large image database retrieval. The image retrieval method based on the color-sift feature dictionary of RGB three channels is adopted. It is a multi-scale image retrieval algorithm, which converts an image into a collection of multiple features, and then calculates the distance between the two image feature vectors. The Euclidean distance is compared to get the results and then realize the image retrieval function. The experimental results show that the algorithm has scale, translation and rotation invariance, and can be well applied. It is also a local feature description operator for scale space, image scaling, and rotating images. It has the characteristics of locality, particularity, quantity and efficiency. The sift feature extraction algorithm can deal with the matching problem in the case of translation, rotation and affine transformation between two images, which improves the accuracy and callback rate of image retrieval.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Specific implementation measures

The invention will be further described below in conjunction with the accompanying drawings.

Example 1

The realization of the image retrieval method based on the weight color-sift feature dictionary of the present invention is with reference to Fig. 1, provides following specific embodiment:

Step 1: Randomly select one image for each type of image in the image database to be retrieved to form the training image database. This example uses the Corel-1000 image database, and it is necessary to retrieve the same type of image in the Corel-1000 image database. The image database includes There are 10 categories of images, and each category includes 100 images. In this example, the value of l is 10, a total of 10 categories, and a total of 100 training images are selected.

Step 2: Randomly select a training image in the training image database for grayscale transformation, process it through a direction-tunable filter, select a two-dimensional Gaussian function as the filter kernel function, and select an appropriate filter sliding window size to obtain each The energy function W _σ (x, y, θ) of each pixel point in 2L directions, through the threshold judgment to extract the significant pixel points of the edge of the image, L represents the number of directions, and the value of L is 6, x and y represent the pixel points σ is the filter scale parameter, σ is 1, θ is the value of the direction, the range is 0～2π, and the interval is π/L. In this example, θ is 0, π/6,... ,11π/6,2π.

Step 3: According to the edge salient pixels of the selected training image, extract the color-sift features of the original training color image in the three channels of red-R, green-G and blue-B respectively, and obtain each edge salient pixel of the image The color-sift features dec _r (e), dec _g (e) and dec _b (e) of the three channels of red-R, green-G and blue-B, e represents the eth edge point of the image, e=1,2,...,E, where E is the total number of all edge salient pixels of the training image.

Step 4: Perform step 2-step 3 for each training image in the training image database to extract the color-sift feature of all edge salient pixels of each image, traverse all images in the training image database, and its features are in three The color channels are dec _r,m (e _m ), dec _g,m (e _m ) and dec _b,m (e _m ), m=1,2,...,M, M is the size of the training image database , em represents the em-th edge pixel of the _m -th training image, _em = 1, 2,..., E _m , and _Em is the total number of all edge salient pixels of the _m -th training image.

Step 5: For the color-sift feature dec _r,m (e _m ) of the red-R channel of all edge significant pixels of all training images, through K-means clustering calculation, take K cluster centers to get red-R The two-dimensional feature dictionary cod _r of row w and column K of the channel, K is the number of K-means clustering centers, that is, the size of the dictionary, the same method is used in the green-G channel and the blue-B channel, respectively. dec _g,m (e _m ) and dec _b,m (e _m ) perform the same calculation of the red-R channel to obtain the two-dimensional feature dictionary cod _g of the green-G channel w row and K column, and the blue-B channel A two-dimensional feature dictionary cod _b with w rows and K columns. In this example, w is a sift feature dimension with a size of 128, and K takes a value of 500, that is, the dictionary size is 500.

Step 6: Input the search image, which is one of the bus images. For the search image, perform steps 2-3 and also extract the red-R, green-G and blue-B color- The sift features sec _r (e), sec _g (e) and sec _b (e), through the feature dictionary cod _r , cod _g and cod _b obtained in step 5, are encoded and calculated to obtain the weight of the retrieved image based on the color-sift feature dictionary Value histogram feature X, weight histogram feature X contains the weight histogram features of red-R, green-G and blue-B three channels;

6a) Retrieve the color-sift feature sec _r (o) of all edge salient pixels of the image on the red channel R, and calculate the Euclidean distance of sec _r (o) corresponding to the K cluster centers of the red channel two-dimensional feature dictionary cod _r , select the clustering center with the smallest distance value as the center of the edge point, and perform first-order distance statistical calculation on all edge significant points to obtain the frequency histogram histogram corresponding to the two-dimensional feature dictionary cod _r of the retrieved image in the red channel _;

6b) Assuming that q _k edge significant points fall in the _k -th cluster center, calculate the centrifugal weight li( k), calculate the maximum centrifugal weight of all the significant points of the cluster center to obtain the weight vector α(k) of the cluster center, and use the following formula to calculate:

$\mathrm{<span\; class="notranslate">\; li</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1,2</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; .</span><span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}}\frac{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; cod</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; cod</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; ,</span>$

α(k)=max(li(u,k)),

Among them, u represents the u-th edge significant point falling in the k-th cluster center, and the value is 1, 2,...,q _k , k represents the k-th cluster center, and the value is 1, 2,. .., K, K represents the number of cluster centers, that is, the size of the dictionary, and q _k represents the total number of edge significant points falling on the kth cluster center. Perform matrix dot multiplication on the frequency histogram his _r of the retrieved image and the corresponding weight vector α(k) to obtain the weight vector hst _r of the retrieved image, that is, multiply the corresponding elements. hst _r is a K-dimensional column vector, and k represents The kth cluster center, the value is 1, 2,..., K, K represents the number of cluster centers, that is, the size of the dictionary;

6c) Do the same calculation as the red channel R in the green channel G and blue channel B, and finally get the weight vector hst _g of the image green channel G and the weight vector hst _b of the blue channel B, for the three channels The weight vector is integrated and calculated to obtain the weight histogram feature X based on the color-sift feature dictionary of the retrieved image.

Step 7: Extract an image to be retrieved from the database of images to be retrieved with a total size of S. Execute steps 2-4 to extract the color-sift feature of all salient pixels on the edge of each image, and then perform step 6 to obtain each An image to be retrieved is based on the weight histogram feature X′ _i of the weight color-sift feature dictionary, traversing all the images in the image database, i=1,2,...,S, S is the total number of images to be retrieved, The database used in this example is Corel-1000, including 10 categories, each category includes 100 images, and the value of S is 1000. .

Step 8: Perform similarity matching based on the weight histogram feature between the retrieved image and the image to be retrieved, and calculate the 2-norm similarity distance to obtain D _i (X, X′ _i ).

Step 9: Arrange each image to be retrieved according to its D _i (X,X′ _i ) value in ascending order, and display the first n images among them as the retrieval result, i=1,2,... ., S, S is the total number of images to be retrieved, n is the number of returned retrieved images, and the value is a positive integer determined by humans. In this example, the value of n is 20. The present invention successfully retrieves 20 bus images related to the bus image from the Corel-1000 image data, but with respect to this item, the retrieval accuracy rate is 100%.

Accurate and fast description of image content has always been the focus and difficulty of image retrieval technology. Traditional image feature extraction methods are basically developed around the image's color, texture, shape and spatial relationship. The present invention first randomly selects the training image in the training image database for grayscale transformation, processes it through a direction-adjustable filter, and performs red-R, green-G and The color-sift features are extracted from the three blue-B channels, and the color-sift features of the red-R, green-G, and blue-B channels of each edge salient pixel point of the training image are obtained. The retrieved image and the image to be retrieved are encoded and calculated based on the feature dictionary to obtain the weight histogram based on the color-sift feature dictionary, and the retrieved image and the image to be retrieved are matched based on the similarity of the weight histogram feature to obtain the retrieval result and improve The efficiency, speed and callback rate of the retrieval process are improved.

Embodiment 2 The image retrieval method based on the weight color-sift feature dictionary is the same as that in Embodiment 1

This example also selects the Corel-1000 image database. The image database includes 10 categories of images, and each category includes 100 images. The same retrieval process as in Embodiment 1 is performed for each image in the database, and the number of retrieved images n is calculated when returning For the average retrieval accuracy rate of each of the 10 categories at 20 o'clock and the average retrieval accuracy rate of 1000 images in all 10 categories, the retrieval results are counted and tabulated, and compared with several well-known retrieval methods in the prior art in the art For example, the method proposed by Jhanwar and Hung was compared with the method based on color-texture-shape, the method based on SIFT-BOF and the method based on SIFT-SPM. The comparison results are shown in Table 1. As can be seen from Table 1, the average retrieval accuracy rate of all 10 categories of 1000 images of the present invention when the number of returned retrieval images n is 20 is significantly higher than each of the above-mentioned retrieval methods for comparison, and each of the 10 categories The average retrieval accuracy of 100 images is higher than most of the retrieval methods used for comparison. Therefore, when the present invention is applied to different types of images for retrieval, it can obtain a higher average retrieval accuracy rate, and is suitable for image retrieval of large-scale image data with many types of images, and can obtain relatively stable results for each type. , Better average retrieval accuracy.

Table 1

The above are two examples of the present invention, which do not constitute any limitation to the present invention. The simulation experiment shows that the present invention can not only improve the speed when applied to large image databases, but also achieve higher accuracy for retrieval results. and callback rate.

In summary, the image retrieval method based on the weighted color-sift feature dictionary of the present invention is mainly dedicated to improving the speed, accuracy and callback rate when the prior art is applied to a large image database. The steps of the method are as follows: randomly select a training image in the image to be retrieved, and perform grayscale transformation on the image; process the training image through a direction-stable filter; combine the results of the direction-stable filter to extract the edge of the image; extract all the training images The color-sift features of all training images are constructed by K-means clustering on the color-sift features of the edge pixels of all training images; input the image to be retrieved and perform the same steps as the training image for the retrieved image and the image to be retrieved Extract the color-sift feature; extract the weight histogram feature based on the color-sift feature dictionary for the retrieved image and the image to be retrieved; perform similarity matching based on the weight histogram feature between the retrieved image and the image to be retrieved in the database; Matching Results displays the image retrieval results. The present invention has the advantages of fast retrieval speed, high accuracy and callback rate especially for large-scale image database retrieval, and can be applied to real-time human-computer interaction and image retrieval of large-scale image databases.

Claims (8)

1. An image retrieval method based on weight color-sift feature dictionary, is characterized in that: it comprises, Randomly select a training image in the image to be retrieved, extract the edge of the training image, and extract the color-sift feature of all training image edge points, and construct a feature dictionary with the color-sift feature; Input the image to be retrieved and extract the color-sift feature of the image to be retrieved and the edge point of the image to be retrieved, based on the feature dictionary to extract the weight histogram feature of the image to be retrieved and the image to be retrieved; Similarity matching based on weight histogram features; Detect whether to traverse all the images to be retrieved in all databases, if yes, display the image retrieval results according to the similarity matching results, if not, perform similarity matching again; The retrieval image encoding is calculated to obtain the weight histogram feature based on the color-sift feature dictionary of the retrieval image, including the following steps: 1) Retrieve the color-sift feature sec _r (o) of all edge salient pixels of the image on the red channel R, and calculate sec _r (o) corresponding to the Euclidean distance of the K cluster centers of the red channel two-dimensional feature dictionary cod _r , select the clustering center with the smallest distance value as the center of the edge point, and perform first-order distance statistical calculation on all edge significant points to obtain the frequency histogram histogram corresponding to the two-dimensional feature dictionary cod _r of the retrieved image in the red channel _; 2) Assuming that there are q _k edge significant points falling in the _k -th cluster center, calculate the centrifugal weight li( k), calculate the maximum centrifugal weight of all significant points in the cluster center to obtain the weight vector α(k) of the cluster center, and make a matrix for the frequency histogram his _r of the retrieved image and the corresponding weight vector α(k) The dot multiplication operation obtains the weight vector hst _r of the retrieved image, that is, the multiplication of corresponding elements, hst _r is a K-dimensional column vector, k represents the k-th cluster center, and the value is 1, 2,...,K, K represents the number of cluster centers, that is, the size of the dictionary; 3) Do the same calculation as the red channel R in the green channel G and blue channel B, and finally get the weight vector hst _g of the image green channel G and the weight vector hst _b of the blue channel B, for the three channels The weight vector is integrated and calculated to obtain the weight histogram feature X based on the color-sift feature dictionary of the retrieved image. 2. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 1, it is characterized in that: randomly selecting training images in the image to be retrieved comprises randomly selecting for each type of image in the image database to be retrieved l images constitute the training image database. 3. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 2, is characterized in that, extracting the edge of described training image comprises: extracting the edge of described training image is included in training image database Randomly select a training image for grayscale transformation, and process it through a direction-tunable filter, select a two-dimensional Gaussian function as the filter kernel function, and obtain the energy function W _σ (x,y , θ), and through the threshold judgment to extract the edge significant pixels of the image, where L represents the number of directions, x and y represent the coordinate values of the pixels, θ is the value of the direction, the range is 0 to 2π, and the interval is π/ L. 4. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 3, is characterized in that, extracting the color-sift feature of all training image edge points comprises: according to the edge salient pixel of selected training image Point pair the original training color image to extract color-sift features in the three channels of red-R, green-G and blue-B respectively, and obtain the red-R, green-G and blue- The color-sift features dec _r (e), dec _g (e) and dec _b (e) of the three channels of B, e represents the e-th edge point of the image, e=1,2,...,E, E is the total number of all edge salient pixels of the image. 5. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 4, is characterized in that, the color-sift feature of described extracting all training image edge points comprises: to each training image database The color-sift feature extraction of all salient pixels on the edge of a training image, traverse all images in the training image database, and its features in the three color channels are dec _r,m (e _m ), dec _g,m (e _m ) and dec _{b, m} (e _m ), m=1, 2,..., M, M is the size of the training image database, e _m represents the em edge pixel of the _{mth training image, e m =} 1,2,...,E _m , where E _m is the total number of all edge salient pixels of the mth training image. 6. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 5, is characterized in that, builds feature dictionary step to comprise red-R channel green- The color-sift features dec _r,m (e _m ), dec _g,m (e _m ) and dec _b,m (e _m ) of the G channel and the blue-B channel are calculated by K-means clustering, and K is taken The two-dimensional feature dictionaries cod _r , cod _g , and cod _b of the red-R channel, green-G channel, and blue-B channel of the w row and K column of the cluster center are obtained, where K is the K-means cluster center number, which is the size of the dictionary. 7. a kind of image retrieval method based on weight value color-sift feature dictionary as claimed in claim 1, is characterized in that: described weight value histogram feature X comprises red-R, green-G and blue-B The three-channel weight histogram feature, the retrieval image and the image to be retrieved are matched based on the weight histogram feature, and the 2-norm similarity distance is calculated to obtain D _i (X,X _i '). 8. A kind of image retrieval method based on weight color-sift feature dictionary according to claim 1, it is characterized in that: described retrieval image coding calculation obtains the weight value histogram based on color-sift feature dictionary of retrieval image Graph feature, wherein step 2) described to clustering to the q _k edge salient points of the kth cluster center calculates the centrifugal weight li(k) of the salient point for the cluster center, and calculates the cluster center The weight vector α(k) of the cluster center is obtained by the largest centrifugal weight of all significant points, which is calculated by the following formula: $\mathrm{<span\; class="notranslate">\; l</span>}\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{{<span\; class="notranslate">\; \&Sigma;</span>}_{\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; ...</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; K</span>}}\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; sec</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; c</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; d</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>{<span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}<span\; class="notranslate">\; ,</span>$ α(k)=max(li(u,k)), Among them, u represents the u-th edge significant point falling in the k-th cluster center, and the value is 1, 2,...,q _k , k represents the k-th cluster center, and the value is 1, 2,. .., K, K represents the number of cluster centers, that is, the size of the dictionary, and q _k represents the total number of edge significant points falling on the kth cluster center.