深度学习驱动的图像检索自编码散列

172 浏览量更新于2024-08-26 收藏 231KB PDF 举报

"用于图像检索的深度自学散列" 在图像检索领域，深度自学散列（DeepSelf-taught Hashing, DSTH）是一种新兴且具有潜力的技术，旨在解决大规模数据集下的高效检索问题。传统的哈希方法通常包括两个主要部分：哈希码生成和哈希函数学习。然而，大多数现有的哈希技术都基于浅层模型，这类模型在挖掘鲁棒视觉特征和学习复杂哈希函数方面存在内在局限。考虑到深度结构，特别是卷积神经网络（CNNs）在提取高级表示方面的优势，该研究提出了一种将深度结构与哈希相结合的框架。DSTH框架利用CNNs的强大能力，能够从原始图像中学习到更丰富的、对光照、姿态变化等具有良好不变性的视觉特征。这些特征对于提高图像检索的准确性至关重要。在DSTH方法中，首先通过预训练的CNN模型对图像进行特征提取，这些特征通常包含多个层次，能够捕获图像的多尺度信息。然后，通过自学习机制，DTH框架可以调整这些特征，使其适应于哈希编码的需求，即在保持信息完整性的同时，将高维特征映射到二进制哈希码。这个过程不仅考虑了特征之间的相似性，还考虑了它们在哈希空间中的分布，以确保相似的图像在哈希码上尽可能接近。此外，DTH还引入了监督学习的元素，利用有标签的数据来指导哈希函数的学习，进一步优化哈希码的生成。这有助于确保哈希码的区分度，使得在检索过程中，即使微小的视觉差异也能被有效地捕捉到。实验结果通常会对比DSTH与其他浅层哈希方法以及基于CNN的哈希方法，展示其在标准图像检索基准数据集上的性能提升。这些比较可能包括平均精度（mAP）、召回率等指标，以证明DTH在大规模图像检索任务中的优越性和实用性。 "用于图像检索的深度自学散列"这篇研究论文探讨了如何利用深度学习的强大力量改进传统的哈希技术，从而实现更准确、更高效的图像检索。通过结合CNNs和自学习策略，DSTH提供了一个强大的工具，对于处理大量图像数据的检索问题具有重要价值。

Deep Self-taught Hashing for Image Retrieval

Ke Zhou

Huazhong University of

Science and Technology

k.zhou@hust.edu.cn

Yu Liu

Huazhong University of

Science and Technology

lightyear416@163.com

Jinkuan Song

University of Trento

jingkuan.song@unitn.it

Linyu Yan

Hubei University of

Technology

yanranyaya@hust.edu.cn

Fuhao Zou

Huazhong University of

Science and Technology

fuhao_zou@hust.edu.cn

Fumin Shen

University of Electronic

Science and Technology of

China

fumin.shen@gmail.com

ABSTRACT

Hashing is a promising technique to tackle the problem of

scalable retrieval, and it generally consists two major com-

ponents, namely hash code generation and hash functions

learning. The majority of existing hashing fall under the

shallow model, which is intrinsically weak on mining ro-

bust visual features and learning complicated hash func-

tions. In view of the superiority of deep structure, espe-

cially the Convolutional Neural Networks (CNNs), on ex-

tracting high level representation, we propose a deep self-

taught hashing (DSTH) framework to combine deep struc-

tures with hashing to improve the retrieval performance by

automatically learning robust visual features and hash func-

tions. By employing CNNs, more robust and discrimina-

tive features of the images can be extracted to beneﬁt the

hash codes generation. Then, we apply CNNs and Multi-

layer Perceptron under deep learning scheme to learn hash

function in supervised process by using the generated hash

codes as labels. The experimental results have shown that

the DSTH is superior to several state-of-the-art algorithms.

Categories and Subject Descriptors

H.3.3 [Information Storage and Retrieval]: Information

Search and Retrieval; I.5.2 [Pattern Recognition]: Design

Methodology; classiﬁer design and evaluation

Keywords

Data Hashing; Deep Learning; Self-taught; Convolutional

Neural Networks

1. INTRODUCTION

Hashing is a promising technique in terms of similarity

search over large scale dataset. Actually, it is a special way

of dimensionality reduction, mapping high dimensional fea-

Permission to make digital or hard copies of all or part of this work for

personal or classroom use is granted without fee provided that copies are not

made or distributed for proﬁt or commercial advantage and that copies bear

this notice and the full citation on the ﬁrst page. Copyrights for components

of this work owned by others than ACM must be honored. Abstracting with

credit is permitted. To copy otherwise, or republish, to post on servers or to

redistribute to lists, requires prior speciﬁc permission and/or a fee. Request

permissions from permissions@acm.org.

MM’15, October 26-30, 2015, Brisbane, Australia.



2015 ACM. ISBN 978-1-4503-3459-4/15/10 ...$15.00.

DOI: http://dx.doi.org/10.1145/2733373.2806320.

ture to compact hash code. Since the Hamming distance

between two binary hash codes can be computed eﬃciently

by using bit XOR operation and counting the number of

non-zero bits, an ordinary PC today would be able to do

millions of Hamming distance computation in just a few

milliseconds. As a result, hashing shows incomparable su-

periority in fast similarity search. Currently, the research

on hashing confronts two challenge: ﬁrst, how to precisely

extract representative image features; second, how to tackle

the problem of semantic gap, leading to accurate transfor-

mation from image feature to hash code.

Aiming at ensuring the eﬀectiveness of hashing, hash code

should preserve the property of discrimination, i.e., objects

withsimilarsemanticshouldbemappedintosimilarhash

codes and vise versa. Several data-aware hashing methods

have been proposed by introducing the machine learning

tricks into the ﬁeld of hashing to enhance the eﬀectiveness

of hash codes [2, 5, 10]. Self-taught hashing (STH) [7] is

proposed and considered as one of state-of-the-art works.

However, it suﬀers from overﬁtting problem since the oper-

ations of generating hash codes for training data and hash

function for testing data are independently handled, which

leads to poor generalization ability. Minimal loss hashing

(MLH) [8] have shown higher search accuracy than unsu-

pervised hashing approaches, but they all impose diﬃcult

optimization and slow training mechanisms. Spectral hash-

ing (SpH) [6] uses a separable Laplacian eigenfunction (LE)

formulation that ends up assigning more bits to directions

along which the data has a greater variance. However, this

approach is somewhat heuristic and relies on an unrealis-

tic assumption that the data is uniformly distributed in a

high-dimensional rectangle. To summarize, the majority of

existing hashing fall under the shallow model, which perform

poor in discovering semantic information. The drawback de-

rives from two aspects: (1) For feature extraction, existing

hashing methods are mainly based on hand-crafted feature,

such as Color Histogram, GIST, SIFT, BoW, etc. However,

those features are limited in aspect of reﬂecting image se-

mantic information, because they represent the semantical

content in just one aspect, either global or local view. (2)

For semantic preservation, shallow model intrinsically could

not explore the high level semantic information contained in

feature data.

To avoid the shortcoming of hashing method of shallow

model, a few deep hashing method have been proposed.

Comparing to shallow learning based hashing, deep learning

下载后可阅读完整内容，剩余3页未读，立即下载

weixin_38725950

粉丝: 3
资源: 901

深度学习驱动的图像检索自编码散列

深度学习驱动的散列检索技术：进展与展望

深度学习图像检索系统Python源码快速检索相似图像

基于Hu不变矩的图像检索技术深度学习实现

用于图像检索的深度渐进式哈希

图像检索-用于图像检索的深度视觉表示的端到端学习算法-附项目源码-优质项目实战.zip

图像检索项目-基于深度学习的图像检索系统python源码.zip

基于深度学习的散列检索技术研究进展.pdf

图像检索图像检索matlab

新型联合稀疏散列法提升图像检索效率

随机投影与散列：隐私保护图像检索新方法

最新资源