The Research Of Region Based Automatic Image Annotation

Automatic image annotation is one of the foundation tasks as well as one of the key technologies in the computer vision. It combines the frontier technologies in many areas, such as image searching, image retrieval and image understanding. In this paper, we make the outdoor static images as the research object. We study the global feature extraction, local feature extraction and semantic mapping, to establish an image annotation model as accurately as possible. Our researches are as follows:1) We propose a Graph based Adaptive Multi-Feature Segmentation Algorithm (GAMS) for the image partition of the low-level features. The segment algorithm we proposed resolved the problems of over-segmentation, non-complete segmentation by using the fusion of combining the texture, spatial location and L*a*b color features. And it also proposed the adaptive selection algorithm of threshold to resolve the problems of the initial threshold selection.2) We propose a new fast local feature description Algorithm, which is called Multi-Resolution Wavelet Transform Descriptor (MRWD). The algorithm not only reduces the instability caused by illumination change, but also maintains a good robustness on the scale and rotation transform compared with the traditional distribution based local feature description algorithm. Meanwhile, MRWD is faster than SIFT in computational speed, and reduces the dimensions. The addition of the local feature can improve the accurate of the image annotation.3) We presents a new framework of image semantic automatic annotation for the problem that the proposed methods were using low-level image features to annotate the images. There are two innovations in the framework we present. Firstly, we construct the characteristic matrix by combining the global features extracted by image segmentation and the local characteristics extracted by the local feature extraction; secondly, this paper establishes the co-occurrence matrix of the semantic and features based the mutual information, using the Priori knowledge. The method improves the robustness of the feature mapping, using the combination of features, and also retains the correlation of the matrix while reducing the data dimensions.