Research Of Image Semantic Automatic Tagging Method

With the popularity of digital camera and other digital imaging products, images are created in a cheaper and easier manner. With the popularity of computer network,microblog, and some image sharing websites, such as Picasa and Flickr, these technologies accelerate the spread of images. It is urgent to advance the research of images so that the capability of images management and understanding could keep consistent with the exploding of images. Automatic image annotation is the core technique for image management and understanding. In this paper, our research is focused on narrowing down the semantic gap, improving the annotation performance and efficiency. The main works and innovative points are as follows:(1)Cross media relevance model in automatic image annotation is studied,and improvement are made to the model:1) In the algorithm, only image segmentation based blob representations used to represent the image, avoiding image segmentation during the automatic annotation process and represent information from different aspects of the image, two representation methods (image rasterization based partition and points of significance detection) are combined to represent the image.2) Co-occurrence relations between label words are considered, the relations are called the subject matter. Multinomial distribution of the label words is used to describe the subject matter. During the image annotation process, probabilities of label words under different subject matters are combined to annotate the testing image; experiments have proven that these improvements have increased the annotation accuracy.(2)A SVM and MBRM fused layered automatic Image annotation method is proposed, first a clustering algorithm is used to cluster all the training images, images which are visually similar are clustered into the same class, and a classifier is build by SVM from each class, then the classifier can be used to add a label to a testing image, the corresponding training set of the label are treated as the similar images of the testing image and these images are used as training images of the MBRM, thus the annotation of the testing image is achieved. During our research of the method, in order to make post annotation results more accurate, some moderations are made to k-means clustering algorithm during the clustering process different influences of the feature of each dimension has on its class are taken into consideration, and different weights are assigned to them accordingly, thus the clustering accuracy is increased subsequent mark, we in the clustering process, improve the k-means clustering algorithm, considering the clustering in every category each dimension characteristics on the influence of this category is different, with different weights, and finally to improve the accuracy of clustering.