The Research On Algorithms Of Relevance Feedback Combining Bayesian And Svm In Image Retrieval

The relevance feedback technique plays an important role in improving the performance of image retrieval. However, the image retrieval process of relevance feedback has many feedback times and bad feedback effect etc. Most of the Bayesian classification algorithm is constrained by the small sample size problem. The SVM and Bayesian classification algorithm is affected by the training sample asymmetry. The solution of these questions plays an important role in improving the retrieval results.In order to solve these above problems, we present a relevance feedback strategy by combining Bayesian and SVM, and propose two improved algorithms on the basis of the algorithm. In this thesis, we study the traditional feedback algorithm based on SVM and Bayesian before the combination algorithm is introduced and do preparation for proposing combination algorithm. The algorithm uses the Bayesian classifier to classify the image library, gather relevant images, remove extraneous images, achieve the compression of the image library, then classified the compressed image library with the SVM classifier. Therefore, the algorithm minifies the searching range and the number of feedback, improves the feedback effect, and solves the problem of feedback of too many times.This algorithm is also helpful to small sample size problem and training sample asymmetry. First, though the training samples of the Bayesian classifier are limited at the beginning of the retrieval, however, the number of samples marked by the user and the number of samples and the training samples increase with the increase of the times of feedback. The effect of Bayesian classifier becomes better. Second, the algorithm parameters of the Bayesian classifier are not fixed; it updates its own parameters with the change of the number of samples. We use different Bayesian classifier in each feedback process, so it will be more reasonable and better. Last, marked by the user in the feedback process is the first K images, only label related images, the rest images are unrelated. We believe that all the unlabeled images in the image library are irrelevant, so it is basically impossible to produce training samples asymmetry.The results show that compared with SVM algorithm, the improved SVM algorithm and Bayesian algorithm, the algorithm presented in this thesis obviously improves the feedback result in the condition of few numbers of feedbacks. The two improved algorithms based on the algorithm are significantly better than this algorithm.