The Study On Region-based Image Annotation Based On LDA Model

With the development of the technologies of internet, photography and multi-media, the number of digital images increase greatly in recent years, and the using of image processing technologies become more and more popular. Thus leads to need of managing and organizing large-scale of image data effectively. Region-based image annotation is an efficient tool to mind the gap between high-level semantics and visual features, which plays an important role in building semantic image retrieval systems. As a result, more and more researchers are interested in it.Traditional annotation methods obtain the region labels by building statistical models of the image regions and visual features directly, which can’t solve the polysemy problem. Some researchers have employed topic models which originally used for text analysis to solve the polysemy problem in recent years. However, they ignore the raw visual features and use quantized features instead, and this reduce the accuracy on two-class annotation problem. Meanwhile, most existing models only take single-cue of features as input, which are not suitable for multi-class image annotation. In this paper, we propose three different annotation models based on LDA model, which can overcome shortcomings mentioned above:1、In LDC model, we consider label of each region is determined by both of its visual features and topic features. The topic features of each region are generated with its label by a supervised mode. As visual features, information of region labels and region topic features are all token into consideration reasonably, LDC can be used in two-class region-based image annotation.2、In the process of optimizing parameters of LDC, the nested iterations are time-consuming. When doing annotation on a large-scale image set, LDC model costs too much time. Thus we propose a light weighted model LDA-C, which utilize LDA model to obtain topic features of regions and use the idea of sample-weighted in LDC to obtain label of regions. Contrast to LDC model, it has a little lower annotation accuracy but reduce the time consumption. So LDA-C can be used on a large-scale image set for two-class annotation.3、Multi-cue s-LDA model is designed for multi-class image annotation. It can take multi-cue of features as its input directly, and each cue of features are quantized independently. Multi-cue s-LDA model utilize multi-cue of visual features to generate region topic features, so these topic features can provide more information than topic features generated with single-cue of visual features.