Research On Image Retrieval Based On Bilinear Mapping And Hash Method

With the development of Internet and multimedia technology, content-based imageretrieval has become a hot research field. This paper introduces low-level feature extraction,dimension reduction, similarity measure of the content-based image retrieval. Aiming atsemantic gap and the curse of dimensionality, we further studied machine learning and imagehashing techniques.Firstly, to make full use of the structure information and fill the semantic gap betweenlow-level features and high-level semantics, a new image retrieval method is introduced,which is based on semi-supervised machine learning and linear mapping. This method takesthe advantage of both labeled and unlabeled samples and establish bilinear mappingmechanism between low-level features and the labels, by constructing regression function,using minimum regression residuals and maintain a smooth manifold of to optimize objectivefunctions.We compare it against the Flexible Manifold Embedding and show a significantimprovement in terms of accuracy and stability based on a subset of the Corel image gallery.Secondly, we introduce the mainstream unsupervised hashing method, improve the hashalgorithm of Locality Sensitive Hashing and Spectral Hashing. We propose a noveldata-dependent projection learning method called bit correlated hashing that each hashfunction is designed to correct the potential errors due to thresholding of the projected data togenerate a bit. Due to thresholding, points may assigned diferent hash bits even though theirprojections are quite close. On the other hand, points that may assigned the same hash biteven though their projected values are quite far apart. To correct these two types of boundary“errors”, we introduce pseudo label and a neighbor-pair set, iteratively to adjust the datacovariance matrix and design regularization term to avoid over-fitting, thus improve the search efficiency. we compare the proposed method against the mainstream unsupervisedhashing method based on a subset of the CIFAR-10and Tiny Images Dataset, includingperformance for different hash code of BCH, comparison of the performance using Hammingranking, performance of PR for different methods and gives a visual diagram of variousmethods.Experimental results show that the proposed method is superior to most mainstreamunsupervised hashing methods in terms of retrieval results.Finally, we made the summary of this article, analysis the shortcomings of the proposedmethod and clear about the direction of further research.