A Subspace Learning Algorithm Based On Multi-view Feature Fusion And Its Application In Image Retrieval

Content-based image retrieval has been an important research topic in the computer vision community,it aims to retrieve query-relevant images from a large-scale dataset.Since visual features are objective descriptions and representations of images,generating visual features of sufficient descriptive and discriminating power plays a crucial role in image retrieval task,and is beneficial for improving the retrieval performance.In recent years,massive important progress has been made in the field of image feature learning.The main research work is divided into two categories,i.e.,traditional methods based on hand-crafted shallow features and deep methods based on deep convolutional neural network.The traditional shallow manual features are derived from locally invariant descriptors and have desirable robustness to the deformation of image cropping,illumination and occluding.By contrast,deep features encode rich high-level semantic information and show unparalleled advantages in many computer vision tasks.Encouraged by the respective success of the two heterogeneous features in image retrieval,researchers attempt to generate more comprehensive and expressive visual features by exploring the complementarity of the two heterogeneous features.It is well known that subspace embedding is an important line of research in multi-view learning,while conventional subspace embedding methods recover a shared subspace from the original multi-view feature spaces.Although existing multi-view learning methods have achieved promising success,these algorithms cannot meet the real-time requirements for retrieval in large-scale scenarios.In addition,existing methods are highly dependent on the visual features of images and are prone to noise interference in complex visual scenes.This paper focuses on the research over multi-view learning in feature learning and representation.In the context of related applications in image retrieval,we study how to further improve the performance of multi-view feature learning and optimize the efficiency of the algorithm.The main work of this paper is summarized as follows:(1)Inspired by Hashing learning,we propose a binary multi-view fusion framework for directly recovering a latent Hamming subspace from the multi-view features.More specifically,the multi-view subspace reconstruction and the binary quantization are integrated in a unified framework so as to minimize the discrepancy between the original multi-view high-dimensional Euclidean space and the resulting compact Hamming subspace.In addition,our method is amenable to efficient iterative optimization for learning a compact similarity-preserving binary code.Since both multi-view reconstruction and binary quantization within our framework are built on efficient7)₂ estimator,our method leads itself to fast iterative optimization.(2)Based on the multi-view hamming subspace learning algorithm,we propose a discriminant multi-view hamming subspace learning model.The image ranking problem is transformed into a classification problem in which the supervised information of partial samples is obtained by the user relevance feedback for supervised learning.The discriminative learning module based on large margin sample separation hyperplane optimization is embedded in the binary multi-view subspace reconstruction model,and the unified discriminative binary multi-view subspace learning framework is established,leading to the generated binary codes with excellent discriminating and descriptive.(3)To further improve the performance of multi-view learning,we propose a unified subspace-based self-weighted multi-view learning framework for instance retrieval.In contrast to the previous methods in which the variances in multi-view features are ignored,we integrate the self-weighted learning mechanism into the multi-view subspace learning framework such that the weights of different views are learned and assigned.