Research On Image Approximate Nearest Neighbor Retrieval Algorithm Based On Hash Algorithm

Image retrieval is a challenging subject in image understanding and computer vision tasks.There is a problem of "semantic gap" in image retrieval.The "semantic gap" is the difference between the image semantics understood by computers and the image semantics perceived by people.Another difficulty is to achieve fast retrieval in large-scale image retrieval.This paper makes full use of image semantic information to solve the problem of semantic gap in image retrieval.And this paper combines hash algorithm to achieve fast retrieval of large-scale image datasets.Firstly,this paper studies an image retrieval algorithm based on image visual and semantic features.This algorithm uses image features combined with hash coding to describe image visual features.At the sime time,it uses the word vector generation model to convert image semantic words into word vectors.Then this article uses word vector sets to describe the semantic features of the image.In order to better combine image visual similarity and semantic similarity for image retrieval.This algorithm establishes a rearranged retrieval matrix.In this paper,the effectiveness of the proposed algorithm is verified through experiments.The retrieved image obtained according to the algorithm not only highlights the consistency of the semantic information of the image but also ensures the similarity of the visual features of the image.Secondly,this paper proposes a Joint learning based deep supervised hashing for large-scale image retrieval.The algorithm integrates the learning of hash functions and image classification into the same end-to-end neural network framework.The weight of the classifier and the hash code corresponding to the entire database can promote each other during the training process.Moreover,the algorithm sets hyperparameters on the traditional paired supervision information to make the output of the neural network closer to the real discrete hash code.The algorithm achieves the best results in the image retrieval experiments on the two datasets of MS-COCO and CIFAR-10.Finally,this paper proposes an improved algorithm for joint learning based deep supervised hashing.In order to reduce the difference between the image representation obtained by network training and the hash code corresponding to the entire database.The algorithm adds a new loss function on the basis of the original algorithm.Simultaneously,to make the loss function converge faster.This paper sets a hyperparameter on the new loss function.It makes the image representation obtained after network training more discrete.The algorithm obtains higher retrieval results than the original algorithm in image retrieval experiments on two commonly used data sets.