Research On Relevance Feedback And Long-term Learning In Content Based 3D Model Retrieval

The result of content based 3D model retrieval is often unsatisfied the user because of the semantic gap between the high-level semantics and the low-level feature. Relevance feedback and long-term learning have great influence on the precision of retrieval result, and thus affect the efficiency of designing and creating 3D models deeply. This dissertation puts all attention on the research of how to improve the precision of retrieval by relevance feedback and long-term learning. The major contributions are presented as follows:1) A parallel relevance feedback method is proposedFirst, the basic idea of similarity field optimization based parallel approach for relevance feedback in 3D model retrieval, in which all variables affecting the similarity field are considered equally, is put forward following a concise survey on existing methods. Second, two strategies are given to simplify this mathematical model:one is similarity field rotation to find and remove less important variables, and the other is dimension reduction based on local linear embedding. Third. translate the mathematical model into a problem of multi-objective optimization under certain constraints. Finally, an algorithm named as fast weighted center particle swarm optimization is proposed and used to solve the mathematical model for parallel relevance feedback in 3D retrieval.2) A partial relevance feedback algorithm is put forwardFirst, a local feature descriptor is put forward, in which principle component analysis is used to normalize the pose of the model, mathematical morphology technology is used to repair the disconnected parts of the silhouettes, and real-time Zernike moment and Fourier feature are used to describe the local featureSecond, a framework and some algorithms are given for partial relevance feedback in 3D retrieval based on the local feature descriptor:decomposing the models into local parts; representing each part by a silhouette for partial retrieval, and improving the result by partial relevance feedback. With this method, the system will be able to catch the user's intentions more accurately by finding out which type of local parts are desired or undesired through marked silhouettes by the user, and thus improve the retrieval result greatly.3) A long-term learning approach is brought forwardFirst, in order to mining semantic relation between the models, all the relevance feedback records in the log file are used as a training data to train a support vector machine based multi-class classifier. This classifier is adopted to divide the model lib into many semantic clusters, which are composed of models with similar meaning. Second, K nearest neighbors of a feedback example existing in multi set are used to determine the closest semantic class, and a new merging algorithm is put forward to prevent the semantic cluster from splitting unlimitedly. Finally, a semantic index based on hash function is brought forward to meet the requirement of real-time response of shape retrieval. This semantic index is able to alter its structure adaptively according to the semantic cluster by putting models similar in the meaning together. With it,3D model retrieval becomes more accurately and more rapidly.Based on the work above, a prototype system named SRS-3D is designed and implemented, with which a set of experiments for testing parallel relevance feedback, partial relevance feedback and long-term learning are conducted. Not only case study but also evaluation of validity and efficiency are given to prove the advantage and feasibility of the methods presented in this dissertation.