Research And Application Of Dynamic Weight Algorithm In Relevance Feedback

With rapid development of multimedia technology and internet technology, enlargement of the image resource can be got and has been used in all works by high-speed storage. It is increasingly important to organize, manage and retrieve the image information resource. In order to improve the use ratio of image, the image retrieval method has become the hot research area at home and abroad. Because of the difficulty in semantic feature extraction, the image retrieval based on low-level feature is a common method at present. But this method can easily form the semantic gap between the low-level feature and the high-level feature. Now one method of solving this problem is relevant feedback. The relevant feedback includes many methods, such as querying vector modification, dynamic weight updating, Bayesian parameter, support vector machine, neural network and so on. Bayesian parameter, support vector machine, neural network methods are actually a machine learning process. These are affected by the sample quantity and the feature dimensions. If the sample quantity is too small, the machine can not get accurate feedback information. If the feature dimension is too high, the machine learning process is time-consuming, so it must be reduced often. However the query vector movement and re-weight updating do not have the problems. Because of all above, we studied both methods, and stressed on the re-weighting updating. Through the experiment analysis, we found that the retrieval sometimes falls in Sub-optimum when retrieving the semantically and visually similar images, with the method of dynamic weight updating. In order to solve the problem, we take two measures: the first one is, in the interactive feedback process, to offer the impact of negative examples to retrieval as interactive formation to the computer that would affect the weight. The second is to use Fisher criteria function to change the weight forcefully, so to withdraw from the local optimal region. Further in the thesis, base on the improving this method, we propose a new one that combines both of the two approaches. In the content-based image retrieval, selecting appropriate image features, taking effective method to extract features, and selecting appropriate feature matching algorithm will all affect the retrieval performance. We select the color and shape features in my thesis, and put forward our own equality quantization method. And we adopt the Euclidean distance matching algorithm.Experimental results show that, under the same experimental conditions, our improved dynamic re-weight method outperforms the traditional updating weight method proposed by the predecessor (Rui method),and its performance is improved. Nevertheless the retrieval performance of our combination method is better than the same and different relevance feedback methods. The study of ours contributes to the research of image retrieval and relevance feedback, and it has certain theory reference value and practical significance.