Research Of Image Retrieval Based On Supervised Hashing

With the rapid development of the Internet and multimedia technology, the global digital image information is growing at an alarming rate. How to quickly and effectively retrieve the right image from huge amounts of images for customers urgently needs to be addressed, rapid and effective image retrieval technology is more and more attention.Content-based image retrieval(CBIR) is a branch of computer vision research, which is focus on image retrieval with large scale. Since CBIR can make image representation with much more semantic information, it has caused wide public concern.Efficient CBIR will produce a lot of super high-dimensional vectors. It leads to high computational overhead and huge storage space, which is so-called“curse of dimensionality”. According to the problems above, we use dimensionality reduction technology to compress high-dimensional vectors into a very low dimension, then establish hash index for them. In addition, there exists ”semantic gap” due to the difference between the human understandings to the images with the machine representation. To reduce“semantic gap”, we join in semantic information with human understandings in training process.After researching a lot of related literature, this paper studied many methods about dig data processing. For the problems of ”semantic gap” and ”dimension disaster”, we also studied many compression technologies for high-dimensional data and took a lot of measures. The contributions of our paper are the following.(1)We propose a general framework of dimensionality reduction for scenarios where data points are represented using histograms and corresponding variants. For the histogram representations of the image, we mainly introduce the intersection kernel as measurement, which makes a good performance in retrieval experiment.(2)Training a mapping function to make a direct relationship between the embedded low-dimensional space and the original high-dimensional space to overcome the defects that traditional MDS does not apply to retrieval applications.(3)Some impressive findings based on exhaustive experiments have been observed by our novel dimensionality reduction algorithm on the histogram representation. We can also find a dimensionality to obtain the optimal retrieval accuracy, which can improve the accuracy compared with the original BOF or SPM.(4)With the human understanding of the image add the method of supervised learning,we make a binaryzation for data embedded in low-dimensional space, which further to raise the discriminative of the image data and reduce the computing and storage overhead.(5)We integrate the dimensionality method and the hash algorithm, which resolves the problem that the dimension reduction cannot be improved furthermore. At the same time, the performance of hash index can be improved.Abundant experiments prove that a dimensionality reduction procedure before hashing the data not only improve the discrimination of the image representations, but also decrease the time and computation of training hashing functions later. Supervised hashing we proposed here can use more correct information from human beings to eliminate the semantic gaps between human subjectivity and the math representation of the multimedia data.