Research On Image Semantic Automatic Marking Based On Multi - Distance Learning

As digital cameras and the social network become more and more prevalent, every day there are tens of thousands of digital images being uploaded and shared through the Internet."A picture paints a thousand words." These massive image data has a vast reservoir of semantic information that worth our effort to excavate. And image annotation is the first step for image data management and utilization. But most of these internet images lack semantic labels, or have the problem of incomplete labels and wrong labels. Since manual annotation costs too much time and effort, Automatic Image Annotation (AIA) has been an appealing research topic for a decade that raises the attention of both the academic and industrial circle.The greatest challenge facing image annotation originates from the so called "semantic gap", namely the mismatch between high level image semantics and low level visual perception. A great deal of research efforts has been devoted to bridge the semantic gap. Most of the recent impressive works of AIA can be roughly categorized into two general classes. The first one is visual feature learning techniques, which focuses on how to combine multiple image features. The second class is the semantic modeling methods, which exploits the correlations between semantic concepts. Both these two kinds of AIA techniques achieve significant improvement in annotation performance comparing to former techniques. Integrating these two paradigms to further enhance the performance of AIA is promising. However, as far as we know, very few works have pursued this issue in a unified framework.In this paper, we propose a novel unified model based on Conditional Random Fields (CRF), the Kernelized Conditional Random Fields (KCRF), which establishes tight interaction between visual features and semantic context. In particular, the CRF model is responsible for semantic context modeling. Kernelized Logistic Regression (KLR) with multiple visual distance learning is embedded into the CRF framework to facilitate multiple visual distances learning. We introduce LI and L2regularization terms into the unified learning process for the distance learning and the parameters penalty respectively.The experiments are conducted on two commonly used benchmarks:Corel5k and TRECVID-2005data sets for evaluation. The experimental results show that, compared with the state-of-the-art methods, the unified model achieves significant improvement on annotation performance. In the experiment we also demonstrate that improvement comes from the integration of semantic modeling and multiple distances learning, not from any one of them. In addition, KCRF shows more robustness with increasing number of various visual features.This paper also realizes an image annotation system. It will carry out parameter estimation on default or user provided training image set and conduct image annotation. The system provides user-friendly interface and the operation is simple. Through interactive demonstration, it allows user to judge the annotation result and mark correct labels.