Research On The Relevance Feedback Based On Log Learning For Image Retrieval

The semantic gap existing in traditional content-based image retrieval makes the results, which obtained by only matching low-level visual features, difficult to meet the actual needs. In order to narrow the semantic gap, relevance feedback method is introduced into image retrieval. The method can learn search intent from the process, that users interact with the retrieval system, and thus correct system search strategy to obtain better search results. However, the traditional relevance feedback methods, that don’t save the semantic evaluation information from users and use it only once in the whole retrieval process, cause the loss of a lot of information. By saving the feedback information from users and the introduction of long-term learning mechanism, the system can further improve retrieval efficiency.In this paper, the log-based relevance feedback is studied. A relevance feedback retrieval algorithm based on the feedback log by the introduction of soft-label SVM using support vector machine as the basic model is proposed. A more effective solution is proposed to ease the problem such as sufficient training data, semantic asymmetry and feedback noise throughout the learning process. Through semantic analysis on the log database and multi-rounds feedback mechanism provided by the system, the training samples are increased in bulk to ease the problem of insufficient sample. In addition, the problem of semantic asymmetry is also reduced through the second expansion of the negative samples to add more Irrelevant semantic to the system, thereby improving the ability of the classifier that classified the relevant or irrelevant semantics. Furthermore, the feedback noise is effectively suppressed by softening training sample label. Finally, another relevance feedback method using comprehensive evaluation is proposed to take full advantage of log database with the rich semantic information. It directly involves the similarity score from the log database regarded as an important part of measuring the similarity between the images in the final similarity score.Finally, a log-based prototype system for image retrieval is developed. Meanwhile, we verify the algorithms proposed in this paper in the system. Experiment results show that the algorithms given in this paper have better retrieval performance.