Research On Key Techniques Of Image Semantic Fusio

With the explosive growth of digital images,higher demand is imposed on automatic understanding of images.Firstly,the intelligent image understanding requires that the semantic gap between image representation and image semantic understanding be bridged as much as possible.Secondly,massive images also demand more efficient image processing.Therefore,efficient and accurate semantic annotation of images is imminent.At present,there are still many problems to be solved in the field of image understanding,which are listed as follows.(1)The efficiency and accuracy of existing image labeling cannot meet the need of massive image processing.The increase of images poses unprecedented challenge to the key technologies of image engineering.The performance of such algorithms as image segmentation,image semantic annotation and image semantic fusion directly affects the accuracy of image detection and image classification.Therefore,how to achieve accurate and efficient image annotation is a challenging research topic.(2)The existing image segmentation ignores the spatial information of image pixels or image sub-blocks,which accounts for low accuracy of image segmentation.As a technology widely used in image segmentation,color histogram lacks spatial information,which hinders the accuracy of image segmentation and image classification.In this paper,color histogram and spatial information improves the accuracy of image segmentation.(3)Ambiguities or redundancies are common in semantic annotations,since different annotation systems tend to provide inconsistent,conflict or incomplete annotations to images.In order to provide complete and accurate semantic annotation for images,it is meaningful to fuse different annotations of the same image.Currently,it is difficult to excavate such deep-level semantics as the scene related semantics and emotional semantics,and this situation leads to the insuperable semantic gap.Image semantic annotation and image semantic fusion,which are designed to eliminate the semantic gap between computer understanding and human needs,become hot topics in the field of image understanding.According to the problems mentioned above,this paper has carried out researches in the following three aspects.(1)The paper proposed a hierarchical semantic annotation model for images.This model makes full use of global features and local features.The process of the model is divided into such two stages as model training phase and image annotation stage.During the model training phase,semantic concept trees are built to establish the correlation between the scene semantics and the visual features.During the image annotation stage,the global features of the image are utilized to find the corresponding semantic concept tree.After that,the local features are used to help the image go from the root node to some leaf node,whose local features are similar to the candidate image.The semantic annotations of all the nodes along the route in the concept tree constitute the semantic annotation of the candidate image.In this paper,a scene related semantic concept tree is constructed to organize the semantics related to the scene.In addition,the semantic annotation algorithm proposed in this paper adds fuzzy clustering to the processing of visual features.The selection of semantic annotation in this algorithm needs to refer to the semantic extraction mechanism of Natural Language Processing,and the annotations provided to the images should be in accordance with the semantic level of the natural language.This paper established semantic concept trees for several common scenes.(2)In this paper,an image segmentation algorithm based on fuzzy clustering and spatial information is proposed.This algorithm combines the spatial pyramid and the color histogram of the image to extract the color histogram at different levels,which makes the image classification and image segmentation more flexible.Moreover,because the spatial pyramid itself contains spatial information of image subblocks at each level,the accuracy of image segmentation is greatly improved.(3)An image fusion algorithm based on semantic similarity and multiple-features is proposed.The algorithm combines the semantic information of two images to get more comprehensive and more accurate image information.The main contributions of this fusion algorithm include the four aspects listed as follows.Firstly,a method of calculating similarity between different semantic concepts is given.Secondly,different visual features are provided with different weights,which reflect the importance of features.Thirdly,different semantic annotations are provided with different weights.Fourthly,fuzzy clustering is adopted in finding the final semantic annotations.According to the experimental results,the following conclusions can be drawn:(1)A stratified semantic annotation model is proposed to make full use of both the global features and local features of the images.The model improves the accuracy of the image semantic annotation and reduces the time complexity.(2)An image segmentation algorithm based on fuzzy clustering and spatial information is proposed.The algorithm combines the color histogram of the image and the spatial pyramid to improve the accuracy of image segmentation.(3)An image fusion algorithm based on semantic similarity and multiple-features is proposed to fuse the semantic information of related images,and gets comprehensive and exact annotation information.