CN102142089A - Semantic binary tree-based image annotation method - Google Patents

Abstract

The invention provides an image labeling method based on a semantic binary tree. Step 1, for the image set of a specific scene, use the image segmentation algorithm to segment the labeled images used for learning to obtain the visual description of the image area; Step 2, construct the visual nearest neighbor graph of all images used for learning; Step 3, Establish the semantic binary tree of the scene according to the nearest neighbor graph in step 2; step 4, find the corresponding position from the root node to the leaf node of the semantic binary tree for the image to be labeled under the scene, and place the node to the root node All annotation words are passed to the image. The present invention aims to establish a semantic binary tree for a labeled image set used for training in a specific scene, so as to improve the accuracy of automatic semantic labeling of images after scene classification using image visual features.

Description

An Image Annotation Method Based on Semantic Binary Tree

technical field

The invention relates to an automatic semantic labeling method for images.

Background technique

As a very valuable image description resource, image annotations better reflect the high-level semantic information of images. How to make full use of the annotation word information of the training image is an important means to improve the accuracy of image annotation. The background of the present invention is to extract the semantic scene of the training image based on the comprehensive utilization of the correlation between image semantics and visual features, and establish a visual model for the training images of different scenes, and finally perform semantic classification on the labeled images according to the visual features.

Contents of the invention

The object of the present invention is to provide an image tagging method based on a semantic binary tree that can improve the tagging accuracy of images to be tagged after scene classification.

The purpose of the present invention is achieved like this:

Step 1. For the image set of a specific scene, the image segmentation algorithm is used to segment the labeled image for learning to obtain the visual description of the image area;

Step 2, constructing visual nearest neighbor graphs of all images used for learning;

Step 3, establishing a semantic binary tree of the scene according to the nearest neighbor graph in step 2;

Step 4, for the image to be labeled in the scene, find the corresponding position from the root node to the leaf node of the semantic binary tree, and transfer all the labeled words from the node to the root node to the image.

The method of constructing the visual nearest neighbor graph of all images used for learning is: the visual distance between images adopts the similarity measure bulldozer distance of multi-region integration matching, the vertices of the graph correspond to each image, and the edges connecting the vertices correspond to the images visual distance between.

The method for establishing a semantic binary tree is as follows: at the root node of the binary tree, all marked images in the scene are collected, and the marked words representing the scene correspond to the semantic representation of the root node, and the nearest neighbor graph in step 2 is normalized and segmented Divide the image into two sets, representing the left subtree and the right subtree of the root node respectively, count a significant label word in the two sets except the label word at the root node, and re-determine according to the label word The attribution of each image; the method of finding the prominent marked words is to count the number of occurrences of each marked word in the collection, and use the marked word with the highest frequency of occurrence as the marked marked word; One marked word as a prominent marked word;

Repeat the above operations on the left subtree and right subtree of the root node until there is only one image or no marked words appearing in the collection, and the bottom leaf nodes correspond to images with less frequent marked words.

The present invention uses marked words and visual information to build a semantic binary tree for marked images of specific scenes, and proposes a specific method for building semantic trees for marked images of specific scenes. The vertices of the tree correspond to the most common marked words in this scene. As the semantic tree grows, the semantics corresponding to each leaf node are cut by branches, and the semantics of the sub-nodes are gradually refined, and the represented marked words are gradually specific, tending to and passing The established semantic binary tree obtains corresponding annotation information from the root of the semantic tree of the scene to the leaf nodes of the image to be annotated for the scene.

The present invention aims to establish a semantic binary tree for a labeled image set used for training in a specific scene, so as to improve the accuracy of automatic semantic labeling of images after scene classification is performed using image visual features.

The invention uses the binary tree with keywords in the nodes for image labeling, and has high practical value. It will be of great help to many CBIR applications, such as Google's image search engine.

Description of drawings

Accompanying drawing is the flowchart of the present invention.

Detailed ways

The present invention is described in more detail below in conjunction with accompanying drawing example:

Step 1. For an image set of a specific scene, an image segmentation algorithm is used to segment the annotated image used for learning to obtain a visual description of the image region.

Step 2, construct visual nearest neighbor graphs of all images used for learning. The visual distance between images uses the similarity measure of multi-region integrated matching, the Earth Mover's Distance (EMD). The vertices of the graph correspond to each image, and the edges connecting the vertices correspond to the visual distance between the images.

Step 3, build a semantic binary tree of the scene according to the nearest neighbor graph in step 2. Methods as below.

The root node of the binary tree collects all the labeled images in the scene, and represents the semantic representation of the labeled words of the scene corresponding to the root node. Use the N-Cut (Normalized Cut) bisection algorithm for the nearest neighbor graph in step 2 to divide the image into two sets, representing the left subtree and right subtree of the root node respectively. Count a significant tag word in the two sets except the tag word at the root node, and re-determine the attribution of each image according to the tag word. The method of finding the marked words is to count the occurrence times of each marked word in the set, and take the marked word with the highest number of occurrences as the significant marked word. If there is more than one marked word with the most frequent occurrences, the marked word with a lower word frequency is taken as a significant marked word.

Repeat the above operations on the left subtree and right subtree of the root node until there is only one image or no marked words appearing in the collection. The leaf nodes at the bottom correspond to images of labeled words that appear less frequently.

Step 4, for the image to be labeled in the scene, find the corresponding position from the root node to the leaf node of the semantic binary tree, and transfer all the labeled words from the node to the root node to the image.

Claims (3)

1. image labeling method based on semantic binary tree is characterized in that:

Step 1 for the image set of special scenes, adopts image segmentation algorithm that the mark image that is used to learn is cut apart, and the vision that obtains image-region is described;

Step 2, the vision arest neighbors figure of all images that is configured to learn;

Step 3, the semantic binary tree of setting up described scene according to the arest neighbors figure in the step 2;

Step 4 to the image to be marked under the described scene, finds the relevant position from the root node of semantic binary tree to leaf node, and this node place is passed to described image to all mark words of root node.

2. the image labeling method based on semantic binary tree according to claim 1, the method that it is characterized in that the vision arest neighbors figure of the described all images that is configured to learn is: the visible sensation distance between image adopts the similarity measure dozer distance of the integrated coupling of multizone, corresponding each width of cloth image in the summit of figure, the visible sensation distance between the limit correspondence image on connection summit.

3. the image labeling method based on semantic binary tree according to claim 1 and 2, the method that it is characterized in that the semantic binary tree of described foundation is: the root node place of binary tree has compiled all the mark images in the scene, represent the semantic expressiveness of the corresponding root node of mark word of described scene, arest neighbors figure in the step 2 is adopted two fens algorithms of normalization cutting, image is divided into two set, represent the left subtree and the right subtree of root node respectively, add up the remarkable mark of except the mark word at root node place word in two set, and redefine the ownership of every width of cloth image by this mark word; The method of seeking remarkable mark word is the occurrence number that respectively marks word in the statistics set, and the mark word that occurrence number is the highest is as significantly marking word; If more than one of the maximum mark word of number of times, the mark word that word frequency is lower is as significantly marking word;

The left subtree of root node and right subtree are repeated aforesaid operations, in having only a sub-picture or set, do not have the mark word that significantly occurs, the leaf node correspondence of bottom the image of the lower mark word of occurrence frequency.

Images