随着网络上图像和视频数据的快速增长,传统图像检索方法已难以高效处理海量数据。在面向大规模图像检索时,特征哈希与深度学习结合的深度哈希技术已成为发展趋势,为全面认识和理解深度哈希图像检索方法,本文对其进行梳理和综述。根据是...随着网络上图像和视频数据的快速增长,传统图像检索方法已难以高效处理海量数据。在面向大规模图像检索时,特征哈希与深度学习结合的深度哈希技术已成为发展趋势,为全面认识和理解深度哈希图像检索方法,本文对其进行梳理和综述。根据是否使用标签信息将深度哈希方法分为无监督、半监督和监督深度哈希方法,根据无监督和半监督深度哈希方法的主要研究点进一步分为基于卷积神经网络(convolutional neural networks,CNN)和基于生成对抗网络(generative adversarial networks,GAN)的无监督/半监督深度哈希方法,根据数据标签信息差异将监督深度哈希方法进一步分为基于三元组和基于成对监督信息的深度哈希方法,根据各种方法使用损失函数的不同对每类方法中一些经典方法的原理及特性进行介绍,对各种方法的优缺点进行分析。通过分析和比较各种深度哈希方法在CIFAR-10和NUS-WIDE数据集上的检索性能,以及深度哈希算法在西安邮电大学图像与信息处理研究所(Center for Image and Information Processing,CIIP)自建的两个特色数据库上的测试结果,对基于深度哈希的检索技术进行总结,分析了深度哈希的检索技术未来的发展前景。监督深度哈希的图像检索方法虽然取得了较高的检索精度。但由于监督深度哈希方法高度依赖数据标签,无监督深度哈希技术更加受到关注。基于深度哈希技术进行图像检索是实现大规模图像数据高效检索的有效方法,但存在亟待攻克的技术难点。针对实际应用需求,关于无监督深度哈希算法的研究仍需要更多关注。展开更多
The easy generation, storage, transmission and reproduction of digital images have caused serious abuse and security problems. Assurance of the rightful ownership, integrity, and authenticity is a major concern to the...The easy generation, storage, transmission and reproduction of digital images have caused serious abuse and security problems. Assurance of the rightful ownership, integrity, and authenticity is a major concern to the academia as well as the industry. On the other hand, efficient search of the huge amount of images has become a great challenge. Image hashing is a technique suitable for use in image authentication and content based image retrieval （CBIR）. In this article, we review some representative image hashing techniques proposed in the recent years, with emphases on how to meet the conflicting requirements of perceptual robustness and security. Following a brief introduction to some earlier methods, we focus on a typical two-stage structure and some geometric-distortion resilient techniques. We then introduce two image hashing approaches developed in our own research, and reveal security problems in some existing methods due to the absence of secret keys in certain stage of the image feature extraction, or availability of a large quantity of images, keys, or the hash function to the adversary. More research efforts are needed in developing truly robust and secure image hashing techniques.展开更多
Deniable authentication protocols allow a sender to authenticate a message for a receiver, in a way which the receiver cannot convince a third party that such authentication ever took place. When we consider an asynch...Deniable authentication protocols allow a sender to authenticate a message for a receiver, in a way which the receiver cannot convince a third party that such authentication ever took place. When we consider an asynchronous multi-party network with open communications and an adversary that can adaptively corrupt as many parties as it wishes, we present a new approach to solve the problem of concurrent deniable authentication within the framework of universally composable （UC） security. We formulate a definition of an ideal functionality for deniable authentication. Our constructions rely on a modification of the verifiably smooth projective hashing （VSPH） with projection key function by trapdoor commitment. Our protocols are forward deniable and UC security against adaptive adversaries in the common reference string model. A new approach implies that security is preserved under concurrent composition of an unbounded number of protocol executions; it implies non-malleability with respect to arbitrary protocols and more. The novelty of our schemes is the use of witness indistinguishable protocols and the security is based on the decisional composite residuosity （DCR） assumption. This new approach is practically relevant as it leads to more efficient protocols and security reductions.展开更多
We present a new method to generate efficient multi-level hashing codes for image retrieval based on the deep siamese conv01utional neural network （DSCNN）. Con- ventional deep hashing methods trade off the capabilit...We present a new method to generate efficient multi-level hashing codes for image retrieval based on the deep siamese conv01utional neural network （DSCNN）. Con- ventional deep hashing methods trade off the capability of capturing highly complex and nonlinear semantic informa- tion of images against very compact hash codes, usually lead- ing to high retrieval efficiency but with deteriorated accuracy. We alleviate the restrictive compactness requirement of hash codes by extending them to a two-level hierarchical coding scheme, in which the first level aims to capture the high-level semantic information extracted by the deep network using a rich encoding strategy, while the subsequent level squeezes them to more global and compact codes. At running time, we adopt an attention-based mechanism to select some of its most essential bits specific to each query image for retrieval instead of using the full hash codes of the first level. The attention-based mechanism is based on the guides of hash codes generated by the second level, taking advantage of both local and global properties of deep features. Experimental re- suits on various popular datasets demonstrate the advantages of the proposed method compared to several state-of-the-art methods.展开更多
In this work, image feature vectors are formed for blocks containing sufficient information, which are selected using a singular-value criterion. When the ratio between the first two SVs axe below a given threshold, t...In this work, image feature vectors are formed for blocks containing sufficient information, which are selected using a singular-value criterion. When the ratio between the first two SVs axe below a given threshold, the block is considered informative. A total of 12 features including statistics of brightness, color components and texture measures are used to form intermediate vectors. Principal component analysis is then performed to reduce the dimension to 6 to give the final feature vectors. Relevance of the constructed feature vectors is demonstrated by experiments in which k-means clustering is used to group the vectors hence the blocks. Blocks falling into the same group show similar visual appearances.展开更多
文摘随着网络上图像和视频数据的快速增长,传统图像检索方法已难以高效处理海量数据。在面向大规模图像检索时,特征哈希与深度学习结合的深度哈希技术已成为发展趋势,为全面认识和理解深度哈希图像检索方法,本文对其进行梳理和综述。根据是否使用标签信息将深度哈希方法分为无监督、半监督和监督深度哈希方法,根据无监督和半监督深度哈希方法的主要研究点进一步分为基于卷积神经网络(convolutional neural networks,CNN)和基于生成对抗网络(generative adversarial networks,GAN)的无监督/半监督深度哈希方法,根据数据标签信息差异将监督深度哈希方法进一步分为基于三元组和基于成对监督信息的深度哈希方法,根据各种方法使用损失函数的不同对每类方法中一些经典方法的原理及特性进行介绍,对各种方法的优缺点进行分析。通过分析和比较各种深度哈希方法在CIFAR-10和NUS-WIDE数据集上的检索性能,以及深度哈希算法在西安邮电大学图像与信息处理研究所(Center for Image and Information Processing,CIIP)自建的两个特色数据库上的测试结果,对基于深度哈希的检索技术进行总结,分析了深度哈希的检索技术未来的发展前景。监督深度哈希的图像检索方法虽然取得了较高的检索精度。但由于监督深度哈希方法高度依赖数据标签,无监督深度哈希技术更加受到关注。基于深度哈希技术进行图像检索是实现大规模图像数据高效检索的有效方法,但存在亟待攻克的技术难点。针对实际应用需求,关于无监督深度哈希算法的研究仍需要更多关注。
基金supported by the National Natural Science Foundation of China（Grant No.60502039）,the Shanghai Rising-Star Program（Grant No.06QA14022）,and the Key project of Shanghai Municipality for Basic Research （Grant No.04JC14037）
文摘The easy generation, storage, transmission and reproduction of digital images have caused serious abuse and security problems. Assurance of the rightful ownership, integrity, and authenticity is a major concern to the academia as well as the industry. On the other hand, efficient search of the huge amount of images has become a great challenge. Image hashing is a technique suitable for use in image authentication and content based image retrieval （CBIR）. In this article, we review some representative image hashing techniques proposed in the recent years, with emphases on how to meet the conflicting requirements of perceptual robustness and security. Following a brief introduction to some earlier methods, we focus on a typical two-stage structure and some geometric-distortion resilient techniques. We then introduce two image hashing approaches developed in our own research, and reveal security problems in some existing methods due to the absence of secret keys in certain stage of the image feature extraction, or availability of a large quantity of images, keys, or the hash function to the adversary. More research efforts are needed in developing truly robust and secure image hashing techniques.
基金the National Natural Science Foundation of China (Grant Nos. 60702059, 60633020 and 60573036)by the MIC of Korea, under the ITRC support program supervised by the IITA (IITA-2006-C1090-0603-0026)
文摘Deniable authentication protocols allow a sender to authenticate a message for a receiver, in a way which the receiver cannot convince a third party that such authentication ever took place. When we consider an asynchronous multi-party network with open communications and an adversary that can adaptively corrupt as many parties as it wishes, we present a new approach to solve the problem of concurrent deniable authentication within the framework of universally composable （UC） security. We formulate a definition of an ideal functionality for deniable authentication. Our constructions rely on a modification of the verifiably smooth projective hashing （VSPH） with projection key function by trapdoor commitment. Our protocols are forward deniable and UC security against adaptive adversaries in the common reference string model. A new approach implies that security is preserved under concurrent composition of an unbounded number of protocol executions; it implies non-malleability with respect to arbitrary protocols and more. The novelty of our schemes is the use of witness indistinguishable protocols and the security is based on the decisional composite residuosity （DCR） assumption. This new approach is practically relevant as it leads to more efficient protocols and security reductions.
基金Supported by the Communication Security Foundation of China under Grant Nos.J641, 0130 (国家通信保密基金)the National Natural Science Foundation of China under Grant No.69931010 (国家自然科学基金)
基金This work was partially supported by the National Natural Science Foundation of China （Grant Nos, 61373060 and 61672280） and Qing Lan Project.
文摘We present a new method to generate efficient multi-level hashing codes for image retrieval based on the deep siamese conv01utional neural network （DSCNN）. Con- ventional deep hashing methods trade off the capability of capturing highly complex and nonlinear semantic informa- tion of images against very compact hash codes, usually lead- ing to high retrieval efficiency but with deteriorated accuracy. We alleviate the restrictive compactness requirement of hash codes by extending them to a two-level hierarchical coding scheme, in which the first level aims to capture the high-level semantic information extracted by the deep network using a rich encoding strategy, while the subsequent level squeezes them to more global and compact codes. At running time, we adopt an attention-based mechanism to select some of its most essential bits specific to each query image for retrieval instead of using the full hash codes of the first level. The attention-based mechanism is based on the guides of hash codes generated by the second level, taking advantage of both local and global properties of deep features. Experimental re- suits on various popular datasets demonstrate the advantages of the proposed method compared to several state-of-the-art methods.
基金Project supported by the National Natural Science Foundation of China （Grant No.60502039）, the Shanghai Rising-Star Program （Grant No.06QA14022）, and the Key Project of Shanghai Municipality for Basic Research （Grant No.04JC14037）
文摘In this work, image feature vectors are formed for blocks containing sufficient information, which are selected using a singular-value criterion. When the ratio between the first two SVs axe below a given threshold, the block is considered informative. A total of 12 features including statistics of brightness, color components and texture measures are used to form intermediate vectors. Principal component analysis is then performed to reduce the dimension to 6 to give the final feature vectors. Relevance of the constructed feature vectors is demonstrated by experiments in which k-means clustering is used to group the vectors hence the blocks. Blocks falling into the same group show similar visual appearances.
