Research On Image Multi-Label Annotation Based On Semantic Analysis

The development of computing technologies, broadband communication networks, and mass storage devices has brought a prosperous increase of digital image and the increasing application demands of effective management and information retrieval. There are three milestones in history since the end of20th century,70s, text-based, content-based and semantic-based image retrieval. Image automatic annotation plays a fatal role of semantic-based image retrieval and becomes one of the hot topics. Ontology is a theory that can be used to express the semantic relationship between concepts so that become an effective method to realize image semantic analysis.In this dissertation, an effective image multi-label semantic automatic annotation based on multi-model is proposed. There are two different models, MFBSA and MNKDA, and RSA, for foreground and background semantic detection respectively, in terms of their distinct characters of semantic and visual features, and then semantic correlation analysis based on LSA algorithm is used to refine the annotation results.Probabilistic latent semantic analysis (PLSA) is one of theories which were introduced into image area from text area, and a new image annotation algorithm based on double layer PLSA is proposed in this paper. Two latent topic spaces were trained by PLSA respectively according to the difference of label and visual information, and then the top model was used to achieve the relationship between these two spaces to detect the concepts in images.To confirm the effectiveness of these two algorithms, the experiments had been implemented on the famous data sets. And the results clearly indicate the performance of the proposed ones over other algorithms.