A Study Of Block-Global Feature Extraction In Semantic Image Annotation

Image Semantic Annotation which is widely used in Image Semantic Retrieval is one of the hot and difficult topics of research in recent years. It associates with Machine Learning,Statistical Model,Information Retrieval and so on. It has been widely used in many fields. For instance, it can be used to label images from big database in order to classify and retrieve images effectively, and it is also applied on Medical image analysis and Crime investigation image analysis.The basic idea of automatic image annotation is to find out the semantic classes contained in a given image. This often requires two procedures, namely feature extraction and mapping. In this paper, Feature Data Optimization method and Block-Global Feature Extraction methods are studied. First, on consideration of human visual perception theories, which demonstrate that global features of a visual object will be perceived before its local features, and Interpretation in terms of a global view tends to conform to the semantic structure of the whole scene, even when it involves some distortion or deletion of few details, two-dimensional principal component analysis (2DPCA) and principal component analysis (PCA) are applied to extract the image block-global features, and comparative studies have been done for the performance of block-global feature extraction methods with widely used local feature extraction method such as scale invariant feature transform (SIFT) method. The results show that block-Global feature not only combines the advantages of local and global features, but also can discover multiple semantic meanings in one image.Otherwise, a novel method is proposed that adopting learning vector quantization (LVQ) technique implemented by self-organizing map (SOM) to select the cluster center vectors as the representative samples for the good performance of Support Vector Machine (SVM) when facing the classification problems for small training sample sets. Experimental results show that the proposed method has a better performance than that without LVQ technique.