Automatic Image Annotation Research Based On Semi-CCA

With the rapid development of information technology and the improving ability of information acquisition and information storage, the method to acquire our interesting information has became one of the most hot research field.In the field of CVPR, image retrieval has also being one of the most important research directions. During the early period, people considered to use'Text Based Image Retrieval'because of the not too much data. This kind of method has the advantage of accuracy and intuition. However, with the sharply increment of information, it is a challenging task to add annotation manually. Besides, it will cause ambiguity for different people has different views on the same image sometimes. So, a new method of Content Based Image Retrieval was proposed and developed quickly. In the paper, with the aid of the canonical correlation analysis tools in statistics, we will get the correlation between annotation vector and low-level character vector. Then, we use k-means clustering toward the new vector projected by Semi-CCA to get several ConceptClass. Now, we continue to segment all of the images in every CC(ConceptClass) to get the segmented regions. We extract low-level character vector of every regions as the input of k-means clustering to require several blobs which has similar character individual. At last, we harvest associated probability distribution list between blobs and key words through Multiple Bernoulli Model.In terms of the un-annotated image, we firstly get the segmentation of image through improved JSEG algorithm. As to each region, we extract the Semi-CCA character projection and find which blob this region belongs to. Combining the associated probability distribution list, we can finally find out the most proper annotation words for un-annotated image.