Research On Svm-based Relevance Feedback Techniques In Cbir

With the wide applications of all kinds of digital medical imaging equipments in the medical field, a large number of digital images are produced in hospital everyday. How to manage images efficiently is the key to realize computer-aided diagnosis and to reach the informationalized hospitals of "paperless and no film". As a research hotspot of the computer image retrieval, content-based image retrieval (CBIR) is the essential approach to realize the management of the digital images, which integrates image content into image retrieval rather than the traditional text retrieval. It is crucial to apply CBIR to the management of the medical images.In CBIR, the low-level visual features (color, texture, shape, etc.) are extracted to represent the images. However, the low level features may not accurately express the high level semantic concepts of images. The retrieval results are often difficult to meet user's requirements just using the low-level image features. To narrow down the semantic gap, the relevance feedback is introduced into CBIR, which refines the retrieval results by learning the feedback from users.Benefit from recent progresses on the relevance feedback, a novel image retrieval algorithm based on SVM is proposed for CBIR relevance feedback in the paper. Firstly after analyzing the disadvantage of traditional SVM relevance feedback algorithm, new sampling algorithms are used for the relevance feedback learning through introducing an active learning scheme into images retrieval. Simultaneously, an improved relevance judgment model is proposed through analyzing the feedback models of users. Then a Bsoft SVM classification algorithm is putted forward combining the improved relevance judgment model. Unlike traditional SVM methods which aim to find a separating hyper-plane with the maximal margin, by introducing the soft mark, the separating hyperplane of Bsoft SVM is defined partial to uncertain points to decrease learning risk and to accelerate the classifier convergence. The large number of experiments shows the approach proposed in the paper has a higher retrieval efficiency compared with traditional SVM. Lastly a retrieval system is designed and realized to verify the validity of algorithms proposed in the thesis for the relevance feedback of CBIR.