Research Of Hashing-based Method For Image Retrieval

With the rapid development of Internet information technology,the increasing number of high-dimensional and massive image data has become more and more demanding on computer processing,which has brought great challenges to related visual work such as image retrieval,classification and object detection.In order to cope with the above problems,many scholars have carried out research work on rapid neighborhood search of massive and high-dimensional data.The traditional linear search method and the tree structure-based retrieval method have good performance when dealing with low-dimensional data,but the accuracy rate decreases rapidly and the retrieval time increases in the face of high-dimensional data,which is difficult to cope with the many problems brought about by“dimension disaster”problem.Hash learning based image retrieval method came into being.Hash learning uses more compact binary code to represent image features,which not only greatly improves the retrieval accuracy and memory usage efficiency,but also greatly shortens the retrieval response time.It can better adapt to the retrieval of massive image data.This paper focuses on the massive image retrieval based on hash learning.For the iterative quantization hash algorithm,the natural matrix structure presented in the high-dimensional image descriptor is not considered.When the visual descriptor is represented by a high-dimensional feature vector and a long binary code is allocated,the projection matrix requires expensive space and time complexity.To solve the problem,a hash image retrieval method based on bilinear iterative quantization is proposed.The method uses a compact bilinear projection instead of a single large projection matrix to map high dimensional data into two smaller projection matrices;then iterative quantization is used to minimize quantization errors and generate efficient hash codes.The method performed comparative experiments on a plurality of published image data sets.It achieves performance comparable to the mainstream eight hashing methods,with faster linear scan time and smaller memory footprint.The results show that the method can reduce the impact of high dimensionality of data,thereby improving the performance of ITQ,and to some extent,it compensates for the high memory and time consumption of hash coding in the existing image hash algorithm.In the current hash algorithm,the content of binary hash coded information is limited,and it is difficult to guarantee the retrieval precision.From the feature level,different features may be mapped to the same hash code,and the Hamming distance is used to measure the image between the images.The similarity is not accurate enough.Aiming at the above problems,an image retrieval method based on adaptive hash learning is proposed.This method adaptively assigns different weights to different code points of the query image,and integrates the existing content-based hash image retrieval algorithm.The adaptive characteristics combine the simple and efficient of the former with the accuracy of the latter,avoiding the influence of all the same code weights on the Hamming distance and improving the retrieval performance.Compared with some current mainstream image hashing algorithms on several public image retrieval databases,the adaptive hash learning image retrieval method proposed in this paper has achieved certain improvements in image retrieval accuracy.