CN107451200B - Retrieval method using random quantized lexical tree and image retrieval method based thereon - Google Patents

Abstract

The invention discloses a retrieval method using a random quantized vocabulary tree and an image retrieval method based on the same, which comprises the following steps: (1) generating a nearest neighbor search tree, taking all the feature vectors of the whole database as root nodes of a first section, and segmenting downwards; (2) in the second stage, k points are randomly selected from the whole database to serve as cluster centers, then each feature vector is distributed to the cluster center closest to the feature vector according to the selected similarity measurement method, the whole database is divided into k subsets, and downward segmentation is continued; (3) in the third level, for each k clusters obtained from the second level, k feature points are randomly selected from the feature vector pool thereof as the cluster centers of the next level. (4) And (6) repeating. The image retrieval method of the invention overcomes the problem that the establishment of the vocabulary tree in the prior art needs a large amount of time, can establish the vocabulary tree in a short time and meets the real-time requirement.

Description

Retrieval method using random quantized lexical tree and image retrieval method based thereon

technical field

The present invention relates to the technical field of image retrieval, in particular to a retrieval method using a random quantized vocabulary tree and an image retrieval method based thereon.

Background technique

In recent years, with the development and popularization of digital technology, especially network technology, the development of Internet of Things and computer information collection software and hardware technology, more and more data are collected and stored, and the speed of quantity collection has far exceeded the traditional method. The speed at which they can be handled, and this trend is becoming more and more obvious. Facebook is the world's leading photo sharing site. As of November 2013, about 350 million photos are uploaded every day, while the photo capacity on Facebook alone has reached 250PB; in terms of digital video, YouTube's statistics in 2013 show that , more than 72 hours of video content are uploaded every minute, there are 4 billion website video playback requests every day, and these data are still increasing significantly. For such huge data resources and the same massive access requirements, how to effectively organize, manage and retrieve large-scale databases has become an urgent problem to be solved.

Traditional text-based image retrieval methods use keywords to annotate images, turning image retrieval into keyword search. Its obvious disadvantage is that neither computer vision nor artificial intelligence technology can automatically perform text annotation on images, and need to rely on manual annotation. Due to the continuous expansion of data scale, the speed of manual annotation is far behind the expansion speed of image data, and due to the subjectivity and imprecision of manual annotation, different people have different understandings of images, resulting in no unified standard for image annotation. To overcome the limitations of text-based image retrieval methods, Content Based-Image Retrieval (CBIR) emerged in the 1990s. Different from traditional retrieval methods, it integrates image understanding technology to provide a method for efficient retrieval from large-capacity image databases according to people's requirements.

The basic idea of content-based image retrieval systems is to analyze the visual features of images and retrieve them in context. Its implementation method is to use an image database to store and manage image data, and then embed the content-based image retrieval technology as the engine of the database into the image database to provide the content-based image retrieval function. In the existing content-based image retrieval system, low-level image information is generally used, including the color, texture, shape of the image and the spatial relationship between them, etc., to calculate the similarity between the query image and the target image, and then according to The size of similarity, that is, the degree of matching between image features, is retrieved. Therefore, firstly, feature extraction is used to convert each image in the image library into a point in the image feature space, that is, the corresponding feature vector, and then the image retrieval is carried out according to the feature vector, thereby transforming the content-based image retrieval into It is the retrieval of feature points in the image feature space.

In the case of a small image database, the most commonly used image feature retrieval method is the sequential scanning method. However, with the continuous development of people's means of obtaining information and the continuous growth of information needs, the scale of image databases is getting larger and larger, and the traditional sequential scanning method has been unable to meet the user's requirements for retrieval time. Therefore, the key to content-based retrieval is to establish an efficient indexing mechanism by effectively organizing the data to rapidly narrow the retrieval range and improve retrieval speed.

In previous related studies, researchers have proposed many data indexing methods for specific application fields. However, when processing high-dimensional data, these data indexing methods are all affected by the "dimension disaster" of high-dimensional space. When the data dimension increases, their retrieval performance degrades to sequential scan, even worse than sequential scan performance. Since feature vectors extracted from original images are usually high-dimensional in CBIR research, the indexing of image feature data is inevitably affected by the "curse of dimensionality". The retrieval method based on lexical tree proposed by Nister and Stewenius showed a good retrieval effect in high-dimensional space, but it took a long time to build a tree in high-dimensional space, and it was difficult to meet the requirements of modern databases for retrieval timeliness. Therefore, according to the high-dimensional characteristics of image feature data, establishing an efficient high-dimensional data indexing mechanism is an important challenge faced by current image retrieval research.

SUMMARY OF THE INVENTION

The present invention provides a retrieval method using a random quantized vocabulary tree and an image retrieval method based thereon, aiming at solving the above-mentioned defects in the prior art.

In order to achieve above-mentioned technical purpose, the present invention adopts following technical scheme:

A retrieval method using randomly quantized lexical trees, including the following steps:

(1) Generate a nearest neighbor search tree, take all the feature vectors of the entire database as the root node of the first section, and divide it downwards;

(2) In the second stage, k points are randomly selected from the entire database as the center of the cluster, and then each feature vector is assigned to the nearest cluster center according to the selected similarity measurement method, and the entire database is divided into For k subsets, continue to subsection down;

(3) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure method assigns each feature vector to its nearest cluster center, thus forming k ² clusters on the third level;

(4) Repeat steps (2) and (3) until all the eigenvectors contained in leaf nodes belong to the same class of objects or the number of eigenvectors contained in leaf nodes is lower than a certain limit; each eigenvector has a Its associated class label.

In step (2), if the distance between the feature vector and the center of two or more clusters is equal, then one cluster is randomly selected.

In step (3), the new feature vector selected from the feature vector pool is assigned to the cluster center closest to it. When reaching the leaf node of the assigned cluster center, if all the feature vectors in the leaf node are If the points have the same class label, assign the associated class label to the new feature vector, and then stop the operation; otherwise, re-search in the assigned cluster, select the feature vector with the shortest distance from the new feature vector in the cluster, and set the The class label associated with that eigenvector is assigned to the new eigenvector, and the computation is then stopped.

An image retrieval method based on a retrieval method of random quantized vocabulary tree, comprising the following steps:

(1) First, the image is divided into several overlapping sub-regions by the overlapping block method;

(2) Combine the feature information blocks of the image with their semantic features; since each extracted feature vector corresponds to a point in the feature space, by performing unguided learning on the feature points in the feature space (that is, clustering in data mining) class), dividing all feature vectors in the feature vector library into multiple patterns, so that there are more similarities between patterns in the same class and greater dissimilarity between patterns in different classes; The feature points are marked by class labels, each class label has specific semantic information, and the image feature information blocks are combined with semantic features to interpret different areas of the image to establish an image knowledge base;

(3) Generate a nearest neighbor search tree, take all image feature vectors in the image knowledge base as the root node of the first section, and divide it downwards;

(4) In the second stage, k points are randomly selected from the image knowledge base as the center of the cluster, and then according to the selected similarity measurement method, each image feature vector is assigned to the nearest cluster center, and the entire The database is divided into k subsets, which continue to be sub-sectioned down;

(5) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure The method assigns each image feature vector to its nearest cluster center, thus forming k ² clusters on the third level;

(6) Repeat steps (2) and (3) until the image feature vectors contained in all leaf nodes belong to the same class of objects or the number of image feature vectors contained in leaf nodes is lower than a certain limit; There is a class label associated with it.

In step (1), the overlapping block method divides an image with a size of height×weight into an N×N window, and shifts the row and column directions by Nhop pixels, and divides it into several overlapping sub-regions. Small enough objects contained in the image can be detected, reducing the size of the square window and increasing the number of patches; by overlapping the subintervals with each other, the color histogram is combined with the spatial distribution of colors.

In step (2), in the process of establishing the image knowledge base, use the Chameleon clustering algorithm and the MST-based clustering algorithm to cluster the color histogram feature vector library of the image, set the class label for the clustering result, and establish the Knowledge base of color histogram features.

Adopting the above technical scheme has the following beneficial effects:

(1) The image retrieval method of the present invention overcomes the problem that the establishment of a vocabulary tree in the prior art requires a large amount of time, and can establish a vocabulary tree in a very short time to meet real-time requirements;

(2) The present invention uses the overlapping block method to refine the image into multiple blocks to extract the color histogram of the image as a feature vector library, effectively combining the color histogram of the image with the color space information, overcoming the prior art. The problem of ignoring the spatial characteristics of color in image feature extraction;

(3) The present invention can extract image features more quickly to meet real-time requirements, and at the same time conduct unguided learning on the feature database, and mark different regions of the scene image through region marking to form a knowledge base.

Description of drawings

Fig. 1 is the schematic diagram of the present invention;

Fig. 2 is the schematic diagram of overlapping block method of the present invention;

Figure 3 is the clustering result of the 10648-dimensional RGB histogram of 21-14;

Fig. 4 is the clustering result of the 5000-dimensional HSV histogram of 24-16;

Fig. 5 is the clustering result of the 5832-dimensional Opponent histogram of 24-16;

Figure 6 is the clustering result of the 10648-dimensional Transformed histogram of 21-14;

Figure 7 is a comparison chart of the accuracy of 21-14 groups of RGB histograms;

Figure 8 is a comparison chart of the accuracy of 24-16 groups of RGB histograms;

Figure 9 is a comparison chart of the accuracy of 27-18 groups of RGB histograms.

specific embodiment

The solution will be further described below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, the retrieval method using random quantized vocabulary tree includes the following steps:

(1) Generate a nearest neighbor search tree, take all the feature vectors of the entire database as the root node of the first section, and divide it downwards;

(2) In the second stage, k points are randomly selected from the entire database as the center of the cluster, and then each feature vector is assigned to the nearest cluster center according to the selected similarity measurement method, and the entire database is divided into For k subsets, continue to subsection down;

(3) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure method assigns each feature vector to its nearest cluster center, thus forming k ² clusters on the third level;

(4) Repeat steps (2) and (3) until all the eigenvectors contained in leaf nodes belong to the same class of objects or the number of eigenvectors contained in leaf nodes is lower than a certain limit; each eigenvector has a Its associated class label.

In step (2), if the distance between the feature vector and the center of two or more clusters is equal, then one cluster is randomly selected.

In step (3), the new feature vector selected from the feature vector pool is assigned to the cluster center closest to it. When reaching the leaf node of the assigned cluster center, if all the feature vectors in the leaf node are If the points have the same class label, assign the associated class label to the new feature vector, and then stop the operation; otherwise, re-search in the assigned cluster, select the feature vector with the shortest distance from the new feature vector in the cluster, and set the The class label associated with that eigenvector is assigned to the new eigenvector, and the computation is then stopped.

For large databases, the retrieval method selection of random quantization tree follows the idea of lexical tree, but an important improvement is made on its basis to generate a nearest neighbor search tree. As shown in Figure 1, given the database, there is only one node in the first level, containing all the feature vectors, which becomes the root node. In the second stage, K points are randomly selected from the entire database as the center of the cluster, and then according to the selected similarity measure method, each feature vector is assigned to the cluster center closest to it, and the entire database is divided into K subsections set. In the third level, for each of the K clusters obtained from the second level, K feature points are randomly selected from their feature vector pools as the cluster centers of the next level, and then the similarity measurement method is used to classify each Each feature vector is assigned to its nearest cluster center, thus forming K clusters on the third level. Continue this process until all leaf nodes contain feature vectors that belong to the same class of objects (ie, the node is pure) or the number of leaf nodes contains feature vectors below a certain limit (eg, 50). Each feature vector has a class label associated with it.

When branching the tree, the distance of each data item to other data items can be updated to a smaller value through a nearest neighbor search tree. This strategy ensures that the closest data points in space are more likely to be assigned to the same partition. However, since any data point in a partition is closest to its own cluster center (and not the nearest neighbor) compared to the centers of other partitions, if the data point is equidistant from the centers of two or more clusters , the data points are randomly selected in a cluster.

Search through a random quantization tree for a given new feature vector. The new feature vector follows a specific path of the random quantization tree, and at each layer calculates the distance from the feature vector to the K cluster centers. The result is the cluster with the closest distance from the new feature vector to the K cluster center points. center point. When reaching a leaf node, if the leaf node is clean (ie, all feature vector points in the leaf node have the same class label), assign the associated class label to the new feature vector, and then stop the operation. Otherwise, do a nearest neighbor search in the vector of related clusters, the result of the search is the feature vector with the shortest distance obtained according to the selected similarity measure method, and assign the class label associated with the feature vector to the new feature vector , and then stop the operation.

An image retrieval method based on a retrieval method of random quantized vocabulary tree, comprising the following steps:

(1) First, the image is divided into several overlapping sub-regions by the overlapping block method;

(2) Combine the feature information blocks of the image with their semantic features; since each extracted feature vector corresponds to a point in the feature space, by performing unguided learning on the feature points in the feature space (that is, clustering in data mining) class), dividing all feature vectors in the feature vector library into multiple patterns, so that there are more similarities between patterns in the same class and greater dissimilarity between patterns in different classes; The feature points are marked by class labels, each class label has specific semantic information, and the image feature information blocks are combined with semantic features to interpret different areas of the image to establish an image knowledge base;

(3) Generate a nearest neighbor search tree, take all image feature vectors in the image knowledge base as the root node of the first section, and divide it downwards;

(4) In the second stage, k points are randomly selected from the image knowledge base as the center of the cluster, and then according to the selected similarity measurement method, each image feature vector is assigned to the nearest cluster center, and the entire The database is divided into k subsets, which continue to be sub-sectioned down;

(5) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure The method assigns each image feature vector to its nearest cluster center, thus forming k ² clusters on the third level;

(6) Repeat steps (2) and (3) until the image feature vectors contained in all leaf nodes belong to the same class of objects or the number of image feature vectors contained in leaf nodes is lower than a certain limit; There is a class label associated with it.

As shown in Figure 2, in step (1), the overlapping block method divides an image with a size of height×weight into an N×N window, and the row and column directions are shifted by Nhop pixels and divided into several overlapping blocks. In order to make small enough objects contained in the image to be detected, the size of the square window is reduced and the number of patches is increased; by overlapping the sub-regions, the color histogram is combined with the spatial distribution of colors. Divide an image of size into squares with overlapping subintervals:

Number of rows: blockrows=(height-N)/Nhop+1

Number of columns: blockcols=(weidth-N)/Nhop+1

Number of block patches: numofSamples=blockrows×blockcols

The image size is height×weight pixels, and the size of the resulting processed image is determined by the size of the patch window. By changing the size of the patch window of the image, the number of generated square patches will also change. As the patch window shrinks, the number of square patches produced by processing the image increases, and vice versa. At this point, in the processed image, each pixel is represented by a color histogram of a square patch. The size of the square window directly affects the resolution of the processed image. Therefore, this restricts the size of the patch square window from being too large, so that a larger number of patches are required to represent the complete image information. That means more patches per image.

Simply dividing the image into blocks to extract its color histogram just divides the image into blocks without any semantic information. Combining the feature information blocks of an image with its semantic features is the purpose of establishing an image knowledge base. Since each extracted feature vector corresponds to a point in the feature space, all feature vectors in the feature vector library are divided into multiple patterns, so that there are more similarities between patterns in the same class and greater dissimilarity between patterns in different classes. The feature points are marked by class labels, each class label has specific semantic information, and the image feature information blocks are combined with semantic features to explain different areas of the image.

In step (2), in the process of establishing the image knowledge base, use the Chameleon clustering algorithm and the MST-based clustering algorithm to cluster the color histogram feature vector library of the image, set the class label for the clustering result, and establish the Knowledge base of color histogram features.

The advantages of the present invention are further demonstrated by means of experiments below.

Clustering of Image Feature Data

The feature vector database is clustered by chameleon clustering algorithm and MST-based clustering algorithm, and the objects in the image are divided into multiple clusters, and the clusters are marked to form a knowledge base. And use color images to represent the clustered database. By comparing the original image set, the MST-based clustering result image, and the Chameleon clustering result image, the feasibility of the image feature extraction method in this paper is verified.

In the experiment, due to the large number of images in the image dataset, we only show the clustering results of some images in the database. At the same time, because the feature database has multiple groups, we separately analyze a group of RGB space, HSV space, Opponent space and Transformed space. The clustering results are displayed.

Table 1 Main scenes and their semantic relationships in the RGB knowledge base based on RGB color histogram

As shown in Figure 3, the figure shows the clustering result of the 10648-dimensional RGB histogram of 21-14.

Table 2 Main scenarios and their semantic relationships in HSV knowledge base based on HSV color histogram

As shown in Figure 4, the figure shows the clustering results of the 5000-dimensional HSV histogram of 24-16.

Table 3 The main scenes and their semantic relationships in the Opponent knowledge base based on the Opponent color histogram

As shown in Figure 5, the figure shows the clustering results of the 5832-dimensional Opponent histogram of 24-16.

Table 4 Main scenes and their semantic relationships in the Transformed knowledge base based on the Transformed color histogram

As shown in Figure 6, the figure is the clustering result of the 10648-dimensional Transformed histogram of 21-14.

Extract feature speed

The image size is 720×1280 pixels, the patch window size is 21×21, and the moving pixel value is set to 14 (represented as 21-14), then the size of the processed image is 50×90 pixels; When the size of the patch window is 24, the moving pixel value is set to 16 (represented as 24-16), then the size of the processed image is ), the size of the processed image is 39×70.

Through a more refined overlapping block method, the image is divided into 4500 blocks, 3476 blocks, and 2730 blocks, and the color histogram of each block is extracted. The previous image segmentation mainly extracts the grayscale color histogram or HSV color histogram of the image. In this paper, the RGB color histogram, HSV color histogram, Opponent color histogram and Transformed color histogram of the image are extracted to establish an image. Eigenvector library.

In the experiment, for 35 images with a size of 720×1280, three overlapping block methods of 21-14, 24-16, 27-18 were used to extract the RGB color histogram of the image with 10000 dimensions, HSV color histogram with 10648 dimensions, Opponent Color histogram with 10000 dimensions, Transformed color histogram with 10000 dimensions, Gabor texture features, and SIFT feature points for extracting the same image.

In this experiment, the extraction time of different features of 35 images was counted, and the average value of each image feature extraction is shown in the following table.

table 5

Table 5 shows that using the same image segmentation method, even if the high-dimensional color histogram of the image is extracted, the extraction speed is significantly better than that of the Gabor texture feature extraction method; at the same time, the time spent to extract the SIFT feature points of the entire image is Much slower than extracting its color histogram after binning. Therefore, the feature extraction method adopted in this paper can quickly extract image features.

Accuracy

Since the nearest neighbor retrieval is that the retrieval object and its nearest neighbor belong to the same class, the retrieval accuracy is the highest. We retrieve the result set with nearest neighbors as the criterion. By comparing the retrieval results of the lexical tree, the retrieval results of the random quantization tree and the nearest neighbor retrieval results, the retrieval accuracy of the two trees is analyzed.

RGB histogram accuracy comparison

First, we target multiple dimensions (64, 125, 216, 512, 1000, 2744, 5832, 10648) of different groups (21-14, 24-16, 27-18) in the RGB color space Dimension), a total of 24 training sets were retrieved, and the comparison results are shown in Table 2, Table 3, and Table 4, where KQtree is a random quantized vocabulary tree, and VTree is a traditional vocabulary tree.

RGB histogram accuracy comparison results: As shown in Figure 7, Figure 8, and Figure 9, the accuracy of the random quantization tree is significantly higher than that of the vocabulary tree.

In Figure 7, the accuracy of random quantization tree is the highest in 1000-dimension, reaching 83.03%, and it is shown that the retrieval results of high-dimension are better than those of low-dimension. In Figure 8, the random quantization tree has the highest accuracy when the dimension is 5832, reaching 86.73%, which also shows that the retrieval result of high dimension is better than that of low dimension. In Figure 9, the random quantization tree has the highest accuracy in 512 dimensions, reaching 85.49%. The retrieval effect of high-dimensional data is better than that of low-dimensional data, but it is not obvious, and the retrieval effect of intermediate dimensions is the best.

Comparing Figure 7, Figure 8, and Figure 9 together, we find that the retrieval effect of RGB color histograms in high-dimensional data space is generally better than that in low-dimensional data space. At the same time, for different square window sizes, the smaller the window is. , the more square patches, the richer the image information, but not the more square patches, the higher the retrieval accuracy of the image, different dimensions have different performance, there is no uniform rule.

Build vocabulary tree time

Randomized quantization tree and lexical tree in RGB histogram in multiple dimensions (64, 125, 216, 512, 1000, 2744, 5832) of different groups (21-14, 24-16, 27-18) dimension, 10648 dimensions), as shown in Table 6, Table 7, and Table 8, in seconds, where KQtree is a random quantized vocabulary tree, and VTree is a traditional vocabulary tree.

Table 6 Window 21×21, shift 14, random quantization vocabulary tree and vocabulary tree running time in different dimensions of RGB histogram

Table 7 Window 24×24, shift 16, random quantization vocabulary tree and vocabulary tree running time in different dimensions of RGB histogram

Table 8 Window 27×27, shift 18, random quantization vocabulary tree and vocabulary tree running time in different dimensions of RGB histogram

In the RGB histogram, the random quantization tree runs significantly faster than the vocabulary tree. As the dimension of the RGB histogram increases, the running time of the lexical tree increases exponentially compared to the running time of the random quantization tree.

Randomized quantization of vocabulary tree and vocabulary tree in Opponent histogram of different groups (21-14, 24-16, 27-18) in multiple dimensions (64-dimensional, 125-dimensional, 216-dimensional, 512-dimensional, 1000-dimensional, 2744-dimensional, 5832 dimensions, 10648 dimensions), as shown in Tables 9, 10, and 11, in seconds.

Table 9 Window 21×21, shift 14, random quantization vocabulary tree and vocabulary tree running time in different dimensions of Opponent histogram

Table 10 Windowed 24×24, Shift 16, Randomized Quantized Vocabulary Tree and Vocabulary Tree Running Time at Different Dimensions of Opponent Histogram

Table 11 Window 27×27, shift 18, random quantization vocabulary tree and vocabulary tree running time in different dimensions of Opponent histogram

In the Opponent histogram, the random quantization tree runs significantly faster than the vocabulary tree. As the dimension of the Opponent histogram increases, the running time of the lexical tree increases exponentially compared to the running time of the random quantization tree.

To sum up, it can be seen that the retrieval method of random quantization vocabulary tree is obviously better than vocabulary tree in terms of time efficiency.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and supplements can be made, and these improvements and supplements should also be considered as It is the protection scope of the present invention.

Claims (1)

1. an image retrieval method based on the retrieval method of random quantization vocabulary tree, is characterized in that, the retrieval method that uses random quantization vocabulary tree comprises the following steps: (1) Generate a nearest neighbor search tree, take all the feature vectors of the entire database as the root node of the first section, and divide it downwards; (2) In the second stage, k points are randomly selected from the entire database as the center of the cluster, and then each feature vector is assigned to the nearest cluster center according to the selected similarity measurement method, and the entire database is divided into For k subsets, continue to segment downwards; if the distance between the feature vector and the center of two or more clusters is equal, a cluster is randomly selected; (3) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure The method assigns each feature vector to its nearest cluster center, thereby forming k ² clusters on the third level; when reaching the leaf node of the assigned cluster center, if all feature vector points in the leaf node have the same class label, assign the associated class label to the new feature vector, and then stop the operation; otherwise, re-search in the assigned cluster, select the feature vector with the shortest distance from the new feature vector in the cluster, and correlate the feature vector The associated class label is assigned to the new feature vector, and then the operation is stopped; (4) Repeat steps (2) and (3) until all the eigenvectors contained in leaf nodes belong to the same class of objects or the number of eigenvectors contained in leaf nodes is lower than a certain limit; each eigenvector has a its associated class label; The image retrieval method includes the following steps: (1) First, the image is divided into several overlapping sub-regions by the overlapping block method; specifically, the overlapping block method divides the image with the size of height×weight into N×N windows, rows and columns. The direction is shifted by Nhop pixels and divided into several overlapping sub-regions. In order to make the small enough objects contained in the image can be detected, the size of the square window is reduced and the number of patches is increased; by overlapping the sub-regions, the color histogram is The graph is combined with the spatial distribution of colors; (2) Combine the feature information blocks of the image with their semantic features; since each extracted feature vector corresponds to a point in the feature space, by performing unguided learning on the feature points in the feature space (that is, clustering in data mining) class), dividing all feature vectors in the feature vector library into multiple patterns, so that there are more similarities between patterns in the same class and greater dissimilarity between patterns in different classes; The feature points are marked by class labels, each class label has specific semantic information, and the image feature information blocks are combined with semantic features to interpret different areas of the image to establish an image knowledge base; specifically, to establish an image knowledge base In the library process, use the Chameleon clustering algorithm and the clustering algorithm based on MST to cluster the color histogram feature vector library of the image, set the class label to the clustering result, and establish a knowledge base based on the color histogram feature; (3) Generate a nearest neighbor search tree, take all image feature vectors in the image knowledge base as the root node of the first section, and divide it downwards; (4) In the second stage, k points are randomly selected from the image knowledge base as the center of the cluster, and then according to the selected similarity measurement method, each image feature vector is assigned to the nearest cluster center, and the entire The database is divided into k subsets, which continue to be sub-sectioned down; (5) In the third level, for each of the k clusters obtained from the second level, randomly select k feature points from their feature vector pools as the cluster centers of the next level, and then use the similarity measure The method assigns each image feature vector to its nearest cluster center, thus forming k ² clusters on the third level; (6) Repeat steps (2) and (3) until the image feature vectors contained in all leaf nodes belong to the same class of objects or the number of image feature vectors contained in leaf nodes is lower than a certain limit; There is a class label associated with it.

Images