Content-based Feature Mixed Image Retrieval

In the real world, the image often uses to describe the truth and store theinformation as a media. It’s different from the traditional characters and numbers.Because the traditional characters and numbers are too simple, they can’t use a fewword to describe a large number of complex semitic information, visual features, thetime and space information.This paper presents the idea of using a feature mixed retrieval model to searchthe image database. At first, this paper introduces the common feature extractionmethods algorithms and performance evaluation. Then this paper studies a content-based image retrieval model based on the random decision trees. First, we use somesquare patches of random sizes to sample at random locations in the image, and resizeby bilinear interpolation to a fixed-size. Second, we propose to build some ensemblesof totally randomized trees for indexing patches as feature vectors. Third, using thelocality-sensitive hashing method searches the nearest neighbors. At last, we calculatethe distance metric between each feature vectors to compute the similarity of twopictures.Base on the model which uses the random decision trees clustering the same typeimages, we do more study on using several features mixed retrieval model. We aim atto solve the problem of curse of dimensionality. First, we add SIFT feature into themodel based on random decision trees. We simulate the camera shaking to remove theunstable feature points, and further use the principal component analysis(PCA)method to reduce the dimension of the index. In the experiment, the results show thatif the SIFT feature detection have a bigger weight, the system will be more accuracy;if the based-on the random decision trees have more weight, the system may have abetter recall.