Research Of Region-based Image Retrieval And Image Annotation

With the rapid development of Internet, the diversification of network platform, and the popularity of digital electronic devices, digital image storage grows explosively. How to quick-ly query and efficiently organize these images becomes an hot issue to be solved.As the most important part of the image content, saliency object plays a key role in image retrieval. In order to effectively use the content of the background information, this paper used the saliency region with the surrounding adjacent valuable background regions to represent image content. In the same time a new matrix descriptor for the similarity measure of back-ground region was proposed. The measure considers not only the similarity of the low-level visual features but also the position relation between regions. We have done some experiments on different databases to verify the effectiveness of the proposed representation method, and completed the contrast experiments with two famous image retrieval algorithms. In addition, we proposed a novel measure method-Mean Label Average Precision(MLAP) considering the image labels of foreground and background. The evaluation results show that the proposed algorithm has obvious superiority.Image region semantic annotation algorithm based on multi-model is to explore the re-lationship between label and image content and mark the label on the corresponding position of image. Firstly, obtaining the visual feature of image region after twice bag-of-words pro-cessing. Secondly, the algorithm applies Conditional Random Field model to detect the labels of image regions, Finally, using Latent Semantic Analysis algorithm to do the error correction according to the region position information. The experiments results indicate the accuracy of region annotation be improved a lot after correction.In this paper, the original Probabilistic Latent Semantic Analysis was improved for image multi-label annotation. By the learning of multi-layer Probabilistic Latent Semantic Analysis, the ’bridge’ between visual features and semantic concepts of image is established, and could be used to predict the semantic concepts of new images. Experiment results show that our automatic image annotation model based on multi-layer Probabilistic Latent Semantic Analysis can achieve promising performance for multi-labeling, and outperform previous methods on standard Corel dataset.