Research Of Hash Image Retrieval Algorithm Based On Improved LLE

With the rapid development of the Internet era,image data has also grown exponentially.The "Dimensional disaster" degenerates traditional image retrieval technology in current image retrieval tasks.In the research of image retrieval,feature extraction and index construction are the key technologies in retrieval tasks.In order to reduce the image dimension,many researchers have turned their attention to manifold learning.Its essence is to discover the inherent low-dimensional manifold structure and embedded mapping relationship of high-dimensional observation data sets.The hash method reduces the space occupied by data storage and speeds up retrieval by transforming the original image into a compact binary code representation.Therefore,the combination of manifold learning and hash function can be more suitable for large-scale image databases.Image retrieval based on hash can be roughly divided into two phases,reducing the dimensionality and then quantizing.The impact of these two stages on the final image retrieval performance is critical.In order to reduce the loss caused by indirect optimization,researchers have considered keeping the similarity between the original data by learning the optimal binary coding directly in the Hamming space.How to better reflect the similarity between the original data is also important for learning the best binary encoding.The main work of this article is as follows:A hash algorithm based on sparse Locality linear embedding is proposed.This algorithm improves the classic algorithm LLE in manifold learning.It uses a sparse weight representation to make each data point find the nearest neighbor that is more suitable for itself to achieve better preservation of the manifold structure between the original data.Learn and preserve manifold structures directly in Hamming space.That is,the binary code is reconstructed from similar data points in the original feature space.This not only keeps the similarity between the original data intact,but also reduces the error caused by indirect optimization.A hash algorithm based on anchor and LLE is proposed.An hash algorithm based on LLE,which is also a common problem in manifold-based hashing methods in the past.In the calculation process,the similarity matrix of the original data needs to be calculated during the learning of the hash function,which will cause the offline learning of the hash function to be very time consuming and take up a lot of memory.In order to further deal with the time and memory consumption issues,an anchor point set is used instead of the entire data set for calculation,that is,by using K-means clustering to generate an asymmetric graph of anchor points to approximate the original similarity matrix in the LLE algorithm,reducing the running time and memory.Tested on multiple commonly used publicly available large-scale image datasets,compared with the current mainstream hashing algorithms,the algorithms proposed in this paper show good retrieval performance.Based on the previous algorithm research,a set of simple and effective image retrieval system was successfully developed,and the effectiveness of the proposed retrieval algorithm was verified by means of corresponding experiments.