Image Retrieval Research Based On Machine Learning

Since image is the basic and widely used form of multimedia information, image retrieval has been a very active research topic in information retrieval. After several decades of development, content-based image retrieval (CBIR) has gained widespread concern. There are three obstacles which hold the application of CBIR:small number training sample problem, imbalance training sample problem and real-time requirements of user-computer interaction.Using similarity measure as entry point, this paper proposed RCAS (Real-time Classification in Asymmetric and Small data collection) algorithm, which combines short-term learning and long-term learning as learning strategy, models both visual similarity and semantic similarity. RCAS can improve the performance of CBIR and proposed a solution for above three problems.Similarity measure is one of the most critical components of CBIR research, and some researches have proved that learning similarity measure from relevance feedback can improve the overall retrieval performance in some degree. First, RCAS learns the semantic similarity from the log data, and then take semantic similarity as prior knowledge, learns visual similarity from the labeled and unlabeled samples. Meanwhile, RCAS employs semi-supervised learning and relevance feedback to deal with the unlabeled samples in the related class and unrelated class. RCAS randomly selects subset of unlabeled data to expand the negative sample set and make use of log data and feedback result to expand the positive sample set, which solves the problem of small sample problem and asymmetry problem. RCAS also return results to user timely in relevance feedback to meet the real-time requirements.In this paper, we implemented RCAS as well as other three algorithms:SVM-AL, Naive-AL and SSAIR, carried out our experiment on expanded COREL database. We evaluated our result on Top N search result accuracy, the PR map and the feedback result. Experimental results show that the RCAS is better than the other three algorithms on various assessment parameters. Finally, we implemented the algorithms and developed a CBIR system named "Thinking&Mining". This system can find related image based on image, and other improved algorithm can be easily added.