RotationInvariantDeepBinaryHashingforFastImageRetrieval资源-CSDN下载

63 浏览量 2021-02-22 00:27:47 上传评论收藏 453KB PDF 举报

这篇研究论文的标题为《Rotation Invariant Deep Binary Hashing for Fast Image Retrieval》，主要内容围绕如何通过深度学习方法，尤其是结合卷积神经网络与哈希技术，来实现快速且旋转不变的图像检索。知识点一：基于内容的图像检索（Content-Based Image Retrieval, CBIR）基于内容的图像检索是指通过分析图像内容来检索出相似或相关图像的技术。这项技术依赖于对图像内容的相似性分析，这对于实现真正的CBIR应用至关重要。知识点二：哈希方法（Hashing Method）哈希方法是一种近似计算与查询图像相关的数据的方法。它在最近几年变得越来越流行，因为它在存储节省和检索效率方面具有显著优势，特别是在处理大规模数据时。哈希码可以通过将图像映射到低维特征空间并进行二值化来获得。知识点三：无监督与监督哈希方法哈希算法大致可以分为无监督和有监督两大类。无监督哈希算法的学习通常与原始输入数据的相似性关系相关。例如，局部敏感哈希（Locality Sensitive Hashing, LSH）使用随机投影将原始数据映射到低维特征空间，然后执行二值化；而谱哈希（Spectral Hashing, SH）则沿着PCA方向使用非线性函数产生紧凑的二进制码。知识点四：旋转不变性（Rotation Invariance）旋转不变性是指图像特征提取或图像处理算法能够识别出图像旋转之后仍然保持不变的特征。在图像检索中，这是非常重要的属性，因为现实世界中的图像可能会在任何方向上被旋转。为了提高检索的准确性，提出的网络需能够处理图像旋转后的数据，提取出与原始图像旋转无关的特征。知识点五：深度二值哈希（Deep Binary Hashing）深度二值哈希是指使用深度学习网络，特别是卷积神经网络（Convolutional Neural Networks, CNNs），来学习紧凑的二值描述符，这些描述符用于图像检索时可以大幅减少存储空间并提升检索效率。提出的方法结合了CNN和哈希技术，通过隐藏层来表征对类别标签起主导作用的潜在概念，并学习出相应的二进制码。知识点六：损失函数的设计为了实现旋转不变性，论文中设计了一种损失函数，旨在最小化描述参照图像和旋转图像之间二进制描述符的差异。这种损失函数有助于训练深度网络，使其能够输出既区分性强又对图像旋转保持不变的二进制码。知识点七：实验结果与对比研究在CIFAR-10和MNIST数据集上进行了实验，结果显示提出的方法非常有效，并取得了最先进的结果。与其他监督方法相比，所提出的网络无需成对输入即可学习二进制码，进一步证明了该方法的优越性。知识点八：旋转不变深度二值哈希网络结构在提出的网络结构中，通过一个隐藏层来学习表示主导类标签的潜在概念的二进制码。这种设计使得网络能够处理旋转后的图像，并且保证了特征描述符的紧凑性和区分性。这篇论文的主要贡献是提出了一个具有旋转不变性的深度二值哈希方法，该方法通过深度学习模型有效学习出对图像旋转不敏感的紧凑二值描述符，并在图像检索任务上取得了优秀的性能。这项技术的应用有望在大规模图像数据处理中提高检索速度和准确率，对于智能搜索引擎、图像监控等领域的实际应用具有重要意义。

资源推荐

资源详情

资源评论

Rotation Invariant Deep Binary Hashing for Fast Image Retrieval

Lai Dai, Jianming Liu, Aiwen Jiang

School of Computer Information and Engineering, Jiangxi Normal University, Nanchang, Jiangxi,

330022

ABSTRACT

In this paper, we study how to compactly represent image’s characteristics for fast image retrieval. We propose

supervised rotation invariant compact discriminative binary descriptors through combining convolutional neural network

with hashing. In the proposed network, binary codes are learned by employing a hidden layer for representing latent

concepts that dominate on class labels. A loss function is proposed to minimize the difference between binary descriptors

that describe reference image and the rotated one. Compared with some other supervised methods, the proposed network

doesn’t have to require pair-wised inputs for binary code learning. Experimental results show that our method is

effective and achieves state-of-the-art results on the CIFAR-10 and MNIST datasets.

Keywords: Rotation Invariant, Deep Binary Hashing, Fast Image Retrieval

1 INTRODUCTION

As is known for us, content-based image retrieval (CBIR) relies on similarity analysis on image content. Fast and

effective comparison is critical to a real CBIR application. Hashing method is one of the popular solutions in recent

years [1, 2, 3, 4] because hashing codes have great advantages on memory saving and retrieval efficiency, especially

when processing large scale data.

There are a lot of hashing algorithms [5, 6] proposed to approximately compute data relevant to the query images. These

methods can be classified into two main categories-unsupervised and supervised methods.

Learning compact binary codes with unsupervised hashing algorithms is usually correlated to the similarity relationship

of the original input data. Locality Sensitive Hashing (LSH) [5] applies random projections to map original data into a

low-dimensional feature space, and then performs a binarization. Spectral Hashing (SH) [6] produces compact binary

codes using non-linear functions along the PCA direction of the given data.

In recent research, many works have shown that supervised information can enhance the performance of binary hash

codes learning because it can take full advantage of label information. However, these methods usually need pair-wised

labels and a large sparse matrix to describe similarities between data points. Krizhevsky et al. [7] proposed a seven-layer

network on ImageNet to extract image features for retrieval, and achieved impressive performance. Banbenko et al. [8]

improved retrieval performance through compressed CNN features. In these methods, the learned CNN features are

directly used to perform pattern matching in Euclidean space, which is definitely inefficient.

There also exist hashing methods combining hashing codes learning with deep architectures. However, most of them are

unsupervised. Xia et al. [9] proposed a supervised hashing approach to learn binary-hashing codes for fast image

retrieval through deep learning. However, matrix-decomposition is needed in their pre-processing stage. When dealing

with large scale data, it requires unaffordable storage and computational cost.

In contrast, the deep network proposed by Lin et al. [10] is more efficient and robust. However, its proposed features are

not rotation invariant. The property of rotation invariant plays great importance in many practical applications. Therefore,

in this paper, we will ameliorate the deep learning convolutional neural networks of Lin et al. [10] to enhance this

property. The method of CNN+Hash proposed by Lin et al. [11] is also efficient, but it also has the same problem with

the paper [10].

In this paper, we proposed a supervised deep binary hashing network. Binary codes are learned by employing some

hidden layer to represent the latent concepts that dominate class labels in data. In order to increase the number of training

samples and the stability of CNN network, we accept to augment data through incorporating rotated version of images.

We propose a loss function called TransformLoss and fine-tune the parameters of the proposed network by minimizing

the distance between binary descriptors that describe the reference image and its rotated one. The proposed deep

Ninth International Conference on Digital Image Processing (ICDIP 2017), edited by Charles M. Falco, Xudong Jiang

Proc. of SPIE Vol. 10420 104201Z-1

Downloaded From: http://proceedings.spiedigitallibrary.org/pdfaccess.ashx?url=/data/conferences/spiep/93262/ on 07/28/2017 Terms of Use: http://spiedigitallibrary.org/ss/termsofuse.aspx

本内容试读结束，登录后可阅读更多

下载后可阅读完整内容，剩余4页未读，立即下载

评论收藏

内容反馈