Research On Cluster-Driven Unsupervised Image Retrieval Methods

As an important task of Internet information processing,image retrieval faces the challenge brought by massive data.The traditional Text-based retrieval method is not applicable to massive image data.To solve this problem,a content-based image retrieval method has been proposed.The content-based image retrieval technology delivers the image content expression and similarity measure to the computer for automatic processing,overcomes the shortcomings of image retrieval using text,and fully exploits the advantages of computer over calculation,greatly improving the retrieval.effectiveness.The hash search method is one of the representative algorithms.It uses the binary code to represent the image.The Hamming distance is used to measure the similarity between images.It has great advantages in algorithm complexity and storage space.In recent years,deep learning has been widely used in image processing tasks,and many deep learning based hashing algorithms have emerged,but most deep hashing algorithms rely on image tags,and tag acquisition is expensive,and most unsupervised.The hash algorithm does not achieve the desired performance.This paper studies the cluster-driven unsupervised image retrieval method,cluster-driven hashing,combines the advantages of hashing method and deep learning,and avoids the cost of labeling image datasets.This unsupervised neural network based unsupervised The Greek method has a wider application prospect.Based on the above questions,the main research contents and results of this paper are as follows:In this paper,a deep embedded hash(UDEH)model is proposed to recursively learn discriminative clustering and generate highly similar binary codes by embedding soft clustering model in depth self-encoder.We use clustered clusters as auxiliary distributions to generate hash codes.By applying binary quantization constraint loss and reconstruction loss of the autoencoder,the algorithm can be jointly optimized by standard stochastic gradient descent(SGD).Then a comprehensive experiment of three popular data sets is carried out,which proves the effectiveness of the algorithm.In this paper,the concept of pseudo-triad is also adopted,and a new method is proposed,namely unsupervised deep triple hash(UDTH),which constructs a triplet self-encoder network and Pseudo-triads are trained as tags.In UDTH,KNN is used to calculate the similarity matrix of all samples,and K most similar images are selected as positive samples,and K most similar images are used as negative samples.With this construction strategy,all images in the dataset are used for triples(as anchors),and the wrong training samples are greatly reduced.In addition,UDTH uses an automatic encoder to keep visual information in the hash code.This paper proposes a deep unsupervised self-evolving hash(DUSH)algorithm to train neural networks from easy to difficult.This model minimizes the problem of parameter selection in previous unsupervised hashing methods,transforming unsupervised hashes into a supervised approach.DUSH is very simple to implement and very efficient.Selecting a pseudo pair in min-batch can greatly reduce the computational complexity.In addition,pseudo-pairs are generated during the training process,which is completely different from the traditional method of generating pseudo-pairs before training.