Research On Large-Scale Image Retrieval Algorithm Based On Hash Coding

Large-scale image retrieval is an important branch in the field of computer vision.In order to ensure retrieval quality and computational efficiency,hash retrieval methods have attracted more and more attention.The idea of the hash retrieval method is to convert highdimensional data into a compact binary hash code and generate similar binary hash codes for similar data items.At present,deep learning combined with hash coding and deep hashing methods achieve outstanding retrieval results in actual retrieval tasks.It is based on powerful image feature learning and hash learning capabilities.In order to learn more effective hash codes to improve retrieval performance,this thesis completes the following work based on existing hash methods.(1)We propose a continuous learning method based on hash center,which improves retrieval performance by reducing quantization loss.Based on the central similarity,this method uses continuous learning to optimize the deep network with symbolic function activation.It solves the ill-posed gradient problem and reduce the feature loss of the data.A large number of comparative experiments are performed on the three benchmark data sets of NUS?WIDE,MS COCO,and CIFAR-10.It shows that the continuous learning method based on the hash center can significantly improve the retrieval performance.Compared with the current classic image retrieval methods,the proposed method can get higher retrieval accuracy.(2)We propose a reordering method based on HashNet.The retrieval result is improved by weighting Hamming distance hash bits,as well as performing fine-grained reordering using the Euclidean distance of pre-relaxed features during the retrieval phase.In large-scale retrieval tasks,the quantization loss is reduced by HashNet.Using the hash bit-weighted Hamming distance for similarity measurement can effectively sort images with the same Hamming distance.In addition,fine-grained reordering of the candidate set obtained by the coarse search using Euclidean distance before relaxation can further improve the search performance.A large number of experiments on the NUS?WIDE,MS COCO and Image Net benchmark datasets show that the HashNet-based reordering method can effectively improve the accuracy and reduce the robustness.