Research On Image Retrieval Based On Cycle Bolcks Color Histogram And LE Algorithm

In image retrieval based on color features, color histogram is one of the most importantdescriptors. The traditional block color histogram can fully express the space information of thecolor, but it loses the advantage of rotational invariance. At the same time, in order to describethe image more completely, the dimensionality of the image features would be very high, whichresult in the increase of storage and computation. And in high-dimensional space, the datapoints are sparse, the Euclidean distance is no longer with the intuitive meaning. This papercombines the content-based image retrieval (CBIR) techniques and manifold learning; firstlywe apply the LE algorithm to image features to get the low-dimensional embedding, and thenretrieve the image in low-dimensional space. The research content and innovations in this paperis as follows:1. This paper introduces the technology of CBIR and manifold learning, and analyses themanifold learning methods and image retrieval methods in detail; then points out the problemsin CBIR and the rationality of the combination of manifold learning and image retrieval.2. Aiming at the shortcomings of losing rotational invariance of the usual blocks colorhistogram, the paper proposes a new strategy--cycle blocks. In this method, each image isdivided into sub-blocks of4×4, and then divides the16sub-blocks into4groups in a circularmanner. For each group, we extract the features separately. And the experiment results showedthat the method we proposed can maintain the rotational invariance of the image better.3. We propose a new LE algorithm based on histogram intersection distance. In thisalgorithm, we use the histogram intersection distance instead of Euclidean distance to computethe k-nearest neighbors, and an image retrieval model based on manifold learning is presented.The experiment’s results shows that compared with the method of direct retrieval on originaldata, the application of LE based on histogram intersection distance improve the retrieval rateas well as retrieval efficiency. During the experiment, the Support Vector Machine, Naive Bayesand KNN are used as retrieval algorithms.