Semantic Segmentation And Annotation Of Image Based On Self-organizing Feature Map (SOFM)

A Semantic Web will be the next stage of World Wide Web's evolution, and the semantic image search is one of the most important parts of the'Semantic Web'. While at the same time the semantic image segmentation and annotation are the basis for semantic image search. However, due to a huge number of the images on the net, the time-consuming of the image processing, it becomes a real challenge for the semantic image segmentation and annotation. Although there have already existed much research on the field of semantic image segmentation and annotation, but most of them are just useful for the special applications, and the segmentation and annotation's efficiency and accuracy are still need to be improved urgently.Most of the previous semantic image annotation systems are the whole image-based ones, they consider the image segmentation and annotation as two irrelevant parts. These systems learn the semantics either from the global image features or from the abstract of the local image features, they are relatively rough and inadequate on the learning of the semantics. There are a small amount of pixel-based semantic image annotation systems which have no problem of adequate semantics learning but are time-consuming of the new image's semantic prediction. In this paper, we propose and construct a semantic segmentation and annotation of image based on self-organizing feature map system after doing much research on the field of the previous work.In this paper, we firstly propose a SOFM(Self-Organizing Feature Map)-based image segmentation algorithm, which not only need not the segmentation number as the input as most of the image segmentation algorithms need to, but also a more accurate segmentation results for the use of the Gauss, Gauss-Laplace, and Gaussian first derivative transforms which generate a 17-dimensional feature vector. The algorithm can also run very fast for the using of the K-Means clustering algorithm.Besides, we considered separately from the image semantic training and prediction. We fully train the images of there every pixels'feature vector, while predict the image segment's semantic by using the center's feature vector. In this way we achieve the quickly unknown image annotation function.Experimental results show that the proposed semantic segmentation and annotation of image based on SOFM system has the properties of high efficiency, high image semantic prediction accuracy and so on. So it is of much practical value of our researches.