一种生成对抗网络用于图像修复的方法

近年来基于深度学习的图像修复方法相比于传统方法,表现出明显优势,前者能更好的生成视觉上合理的图像结构和纹理.但现有的标准卷积神经网络方法,通常会造成颜色差异过大和图像纹理缺失与失真的问题.本文提出了一种新型图像修复深度网络模型,该模型由两个相互独立的生成对抗式网络模块组成.其中,图像修复网络模块旨在解决图像缺失区域的修复问题,其生成器基于部分卷积网络;图像优化网络模块旨在解决修复后图像存在局部色差的问题,其生成器基于深度残差网络.通过两个网络模块的协同作用,图像的视觉效果与图像质量得到提高.与其他先进方法进行定性和定量比较的实验结果表明,本文提出的方法在图像修复质量上表现更好.     

基于金字塔池化网络的道路场景深度估计方法

针对从单目视觉图像中估计深度信息时存在的预测精度不够准确的问题,该文提出一种基于金字塔池化网络的道路场景深度估计方法。该方法利用4个残差网络块的组合提取道路场景图像特征,然后通过上采样将特征图逐渐恢复到原始图像尺寸,多个残差网络块的加入增加网络模型的深度;考虑到上采样过程中不同尺度信息的多样性,将提取特征过程中各种尺寸的特征图与上采样过程中相同尺寸的特征图进行融合,从而提高深度估计的精确度。此外,对4个残差网络块提取的高级特征采用金字塔池化网络块进行场景解析,最后将金字塔池化网络块输出的特征图恢复到原始图像尺寸并与上采样模块的输出一同输入预测层。通过在KITTI数据集上进行实验,结果表明该文所提的基于金字塔池化网络的道路场景深度估计方法优于现有的估计方法。     

基于深度多级小波变换的图像盲去模糊算法

近年来卷积神经网络广泛应用于单幅图像去模糊问题,卷积神经网络的感受野大小、网络深度等会影响图像去模糊算法性能.为了增大感受野以提高图像去模糊算法的性能,该文提出一种基于深度多级小波变换的图像盲去模糊算法.将小波变换嵌入编-解码结构中,在增大感受野的同时加强图像特征的稀疏性.为在小波域重构高质量图像,该文利用多尺度扩张稠密块提取图像的多尺度信息,同时引入特征融合块以自适应地融合编-解码之间的特征.此外,由于小波域和空间域对图像信息的表示存在差异,为融合这些不同的特征表示,该文利用空间域重建模块在空间域进一步提高重构图像的质量.实验结果表明该文方法在结构相似度(SSIM)和峰值信噪比(PSNR)上具有更好的性能,而且在真实模糊图像上具有更好的视觉效果.     

基于深度多级小波变换的图像盲去模糊算法

近年来卷积神经网络广泛应用于单幅图像去模糊问题,卷积神经网络的感受野大小、网络深度等会影响图像去模糊算法性能。为了增大感受野以提高图像去模糊算法的性能,该文提出一种基于深度多级小波变换的图像盲去模糊算法。将小波变换嵌入编-解码结构中,在增大感受野的同时加强图像特征的稀疏性。为在小波域重构高质量图像,该文利用多尺度扩张稠密块提取图像的多尺度信息,同时引入特征融合块以自适应地融合编-解码之间的特征。此外,由于小波域和空间域对图像信息的表示存在差异,为融合这些不同的特征表示,该文利用空间域重建模块在空间域进一步提高重构图像的质量。实验结果表明该文方法在结构相似度(SSIM)和峰值信噪比(PSNR)上具有更好的性能,而且在真实模糊图像上具有更好的视觉效果。     

面向压缩图像复原的网络增强训练方法

高质量的数据是深度卷积神经网络成功的关键因素之一。在计算机视觉领域,常用图像数据集通常以JPEG格式存储。这种有损压缩技术不可避免地会导致原始数据信息的丢失,进而造成利用压缩数据训练的卷积神经网络的性能降低。因此,为了增强卷积神经网络的性能,本文提出了一种面向压缩图像复原的增强训练方法,通过复原压缩图像实现卷积神经网络的性能增强。该方法具体为一个包含复原模块和任务模块的联合增强框架。复原模块致力于恢复有损压缩技术造成的信息丢失;任务模块专注于基于任务需求增强压缩图像。两个模块联合训练,使得压缩图像的复原增强更具有目的性。本文通过图像分类任务的实验表明,与压缩图像相比,该方法能有效地复原压缩图像,增强卷积神经网络的性能。此外,该方法中两个模块间的低耦合性和可替代性保证了该方法的适用性。      

基于深度卷积神经网络的协作频谱感知方法

针对传统卷积神经网络(CNN)频谱感知方法提取特征能力受限于网络结构简单,增加网络结构又容易出现梯度消失等问题,该文通过在传统卷积神经网络中添加捷径连接,实现输入层恒等映射更深的网络,提出一种基于深度卷积神经网络(DCNN)的协作频谱感知方法。该方法将频谱感知问题转化为图像二分类问题,对正交相移键控(QPSK)信号的协方差矩阵进行归一化灰度处理,并作为深度卷积神经网络的输入,通过残差学习训练深度卷积神经网络模型,提取2维灰度图像的深层特征,将测试数据输入到训练好的模型中,完成基于图像分类的频谱感知。实验结果表明:与传统的频谱感知方法相比,在低信噪比(SNR)下、多用户协作感知时,所提方法具有更高的检测概率和更低的虚警概率。     

基于深度卷积神经网络的协作频谱感知方法

针对传统卷积神经网络(CNN)频谱感知方法提取特征能力受限于网络结构简单,增加网络结构又容易出现梯度消失等问题,该文通过在传统卷积神经网络中添加捷径连接,实现输入层恒等映射更深的网络,提出一种基于深度卷积神经网络(DCNN)的协作频谱感知方法.该方法将频谱感知问题转化为图像二分类问题,对正交相移键控(QPSK)信号的协方差矩阵进行归一化灰度处理,并作为深度卷积神经网络的输入,通过残差学习训练深度卷积神经网络模型,提取2维灰度图像的深层特征,将测试数据输入到训练好的模型中,完成基于图像分类的频谱感知.实验结果表明:与传统的频谱感知方法相比,在低信噪比(SNR)下、多用户协作感知时,所提方法具有更高的检测概率和更低的虚警概率.     

基于卷积循环网络与非局部模块的语音增强方法

现有的深度神经网络语音增强方法忽视了相位谱学习的重要性,从而造成增强语音质量不理想。针对这一问题,文中提出了一种基于卷积循环网络与非局部模块的语音增强方法。通过设计一种编解码网络,将语音信号的时域表示作为编码端的输入进行深层特征提取,从而充分利用语音信号的幅值信息以及相位信息。在编码端和解码端的卷积层中加入非局部模块,在提取语音序列关键特征的同时,抑制无用特征,并引入门控循环单元网络捕捉语音序列间的时序相关性信息。在ST-CMDS中文语音数据集上实验结果表明,与未处理的含噪语音相比,使用文中方法生成的增强语音质量和可懂度平均提升了61%和7.93%。     

基于深度学习特征表达的车辆检测和分析

在基于单目视觉的辅助驾驶中,对车载摄像头拍摄的视频进行车辆检测、识别、分析,可以提取出有效信息来提醒司机或控制车辆的行驶,是机器视觉技术挑战问题。该文利用深度特征表达对车载视频进行车辆检测和分析,首先,针对现有卷积神经网络对超清分辨率车载视频分析效果差的问题,提出随机失活池化降维方法改进设计卷积神经网络适应高分辨率视频;其二,针对检测标识的车辆提取行驶状态信息的问题,该文利用现有卷积神经网络的重新训练分析出车辆的行驶方向:前向(F-direction)行驶车辆和对向(R-direction)行驶车辆。实验证明,该文的方法能够实时、有效地检测车辆和分析状态。     

基于空间注意力的图像分类网络研究

针对以往的图像分类方法利用手工提取的特征（或通过神经网络提取的特征）、空间信息关注不足等问题,文章提出一种基于空间注意力的图像分类网络。该网络利用空间注意力模块,对深度网络提取的视觉特征进行空间约束。利用特征的空间信息,使得网络能够对特征在空间上的重要性加以区分,从而使其更具判别性。采用CIFAR-10和CIFAR-100测试集分别进行测试,测试结果表明,该文提出的图像分类网络的图像分类效果明显优于其他深度学习方法。     

Hiding secret data adaptively in vector quantisation index tables

Digital data hiding applications have been developed mainly with a view for sending secret information safely through the Internet or other computer networks. Most of the existing vector quantisation (VQ)-based data hiding methods, however, are capable of embedding only a limited quantity of secret data into the medium, usually a secret bit in every block at most. To improve the embedding capacity without seriously sacrificing qualities of carrier media, an adaptive VQ-based data hiding scheme based on a codeword clustering technique is presented. After the adaptive clustering process, the proposed scheme embeds secret data into the VQ-index table by performing codeword-order-cycle permutation. With the help of the cycle technique, more options can be offered for proper codeword substitution, which adds more flexibility to the whole data hiding scheme and therefore improves the quality of the embedded image. As the experimental results demonstrate, the proposed scheme is capable of providing better image quality and embedding capacity than the currently existing VQ-based hiding schemes. In fact, the stego-image quality is so good that it is visually indistinguishable from the VQ-compressed image     

基于多尺度特征结合细节恢复的单幅图像去雾方法

为提高单幅图像去雾方法的准确性及其去雾结果的细节可见性,该文提出一种基于多尺度特征结合细节恢复的单幅图像去雾方法。首先,根据雾在图像中的分布特性及成像原理,设计多尺度特征提取模块及多尺度特征融合模块,从而有效提取有雾图像中与雾相关的多尺度特征并进行非线性加权融合。其次,构造基于所设计多尺度特征提取模块和多尺度特征融合模块的端到端去雾网络,并利用该网络获得初步去雾结果。再次,构造基于图像分块的细节恢复网络以提取细节信息。最后,将细节恢复网络提取出的细节信息与去雾网络得到的初步去雾结果融合得到最终清晰的去雾图像,实现对去雾后图像视觉效果的增强。实验结果表明,与已有代表性的图像去雾方法相比,所提方法能够对合成图像及真实图像中的雾进行有效去除,且去雾结果细节信息保留完整。     

基于块容量标签的密文图像可逆数据隐藏方法

密文图像的可逆数据隐藏技术既能保证载体内容不被泄露,又能传递附加信息。本文提出了一种基于块容量标签(block capacity label, BCL)的高容量密文图像可逆数据隐藏算法。该方案在图像加密之前进行预处理,首先将图像分为两个区域：参考像素区域和预测像素区域。然后将预测像素区域分为不重叠的块,根据所提出的算法确定分块的BCL,在对图像进行加密之后嵌入BCL,生成加密图像;在秘密数据嵌入阶段,根据BCL和数据隐藏密钥嵌入秘密数据。实验测试了BOWS-2数据集,平均嵌入容量为3.806 8 bpp,与现有方法相比,该方法可以获得更高的秘密数据嵌入容量,并可以实现原始图像的完美重建。     

Wavelet bit-plane based data hiding for compressed images

In this paper, a wavelet bit-plane based data hiding for compressed images is proposed. Image compression not only reduces storage but also benefits transmission. Currently, image encoders including JPEG2000, SPIHT, EZW, etc. also provide multi-stage encoding/decoding. In this paper, the bit-planes of DWT coefficients are selected to carry the secret image according to the multi-stage encoding. The hidden secret image can be extracted progressively from multi-stage decoded images.Experimental results showed that the secret image was embedded and extracted in multi-stage coded images. Furthermore, the structure of secret image could be identified in earlier decoding stages and then refined in later stages. Accordingly, the progressive secret revealing is achieved. In comparison with the other similar schemes, the proposed method achieves the better quality of stego-image than the other two when the hiding capacity is the same.     

A secure high capacity full-gray-scale-level multi-image information hiding and secret image authentication scheme via Tchebichef moments

The growth of image processing tools and applications has made it easy for multi-media content such as music, audio, and video to be manipulated or forged during transmission over the Internet. Efforts, such as information hiding in steganography, have been unable to secure data transmission and prevent its manipulation. Usage of coding theory, including cryptography, is not full proof in the sense that an unauthorized intruder may inject (tampering) and incorporate unintended data to the messages, which can tamper the transmitted data. There is a need for more transparent message information hiding schemes along with information content verification and authentication, as well as accurate tampering detection. In particular, as it is well known, in many current steganography methods, widely used for image information hiding, there are various technical challenges associated with hiding large amounts of image information in images. Some of these challenges relate to which locations, in a given carrier image, information has to be hidden in order to guarantee transparency of the resulting watermarked images, to the ability to extract hidden information accurately, to the performance of hidden secret information authentication and verification at the receiving end, to the dependency of the hidden information on a given carrier image, to the robustness of information hiding schemes to affine transformations such as rotation, and to the amount of data and number of full-scale images one can embed in a given single image carrier. Additionally, as it is well known, many of the existing stenography methods are based on the Discrete Fourier Transform (DFT), the Discrete Cosine Directors (DCT), or the Discrete Wavelength Transform (DWT) methods, which result in high Bit Error Rate (BER) of the extracted data. In this paper we present a secure high capacity image information hiding scheme where two full separate arbitrary full-scale gray level images (versus binary), one hidden information image and one authentication watermark image are hidden/embedded in the Tchebichef moments of a carrier image with very high imperceptibility. Here the second watermark image is used for identification and content integrity verification and authentication of the hidden secret image. Both the hidden secret hidden image and the authentication watermark image are of the same size as that of a given arbitrary carrier image. In particular, with the cost of computer memory getting lower and the bandwidth of transmission channels getting larger, we show how three different watermarked images, but the same to a naked eye, are produced and transmitted to achieve the desired advantages of high accuracy, security, authentication and verification of the recovered information. To the best of our knowledge, this two-full-scale gray images data hiding and hidden secret image information verification and authentication method is the first attempt of its sort. We show here the robustness of the proposed scheme to affine transformations such as rotation, scaling, and translation, the proposed scheme's high image malicious tampering detection and tampering localization and its high quality extracted recovered and authenticated hidden secret images. Additionally, in order to as much as possible keep the integrity of the received information, when watermarked images are rotated during transmission, a new image rotation estimation and recovery algorithm is presented as part of the proposed information hiding scheme. We show the effect of intended tampering attacks namely, cropping, noise, low-pass and high-pass filtering on the presented scheme. We also show how the extracted information accuracy is generally independent of the carrier image, and we present a mathematical analysis for characterizing the conditions under which transparency of the hidden embedded information is generally achieved for any given arbitrary carrier image. The case of how to extract the hidden information when one or two of the watermarked images is (are) lost is also tackled. Finally, experimental results on real images are presented to illustrate the efficiency and capabilities of the proposed method.     

基于多级特征补偿的遥感图像时空融合方法

在许多地球科学应用中要用到大量的高时空分辨力的地球观测数据。时空图像融合方法为产生高时空分辨力的数据提供了一种可行且经济的解决方案。然而,现有的一些基于学习的方法对于图像深层特征提取能力较弱,对于高分辨力图像细节特征利用度不够。针对这些问题,提出一种基于多级特征补偿的遥感图像时空融合方法。该方法使用2个分支进行多层级的特征补偿,并提出了融合通道注意力机制的残差模块作为网络的基本组成单元,可以将高分辨力输入图像的深层特征更为详尽地提取利用。提出一种基于拉普拉斯算子的边缘损失,在节省预训练计算开销的同时取得了很好的融合效果。使用从山东和广东2个地区采集的Landsat和中分辨力成像光谱仪(MODIS)卫星图像对所提出的方法进行实验评估。实验结果表明,提出的方法在视觉外观和客观指标方面都具有更高质量。     

Enhanced hybrid image security algorithms for high definition images in multiple applications

Protection of multimedia information from different types of attackers has become important for people and governments. A high definition image has a large amount of data, and thus, keeping it secret is difficult. Another challenge that security algorithms must face with respect to high definition images in medical and remote sensing applications is pattern appearances, which results from existing regions with high density in the same color, such as background regions. An encryption and hiding based new hybrid image security systems are proposed in this paper for the purpose of keeping high definition images secret. First, one hiding method and two encryption methods are used in two hybrid algorithms. The new hiding algorithm proposed here starts by applying reordering and scrambling operations to the six Most Significant Bit planes of the secret image, and then, it hides them in an unknown scene cover image using adding or subtracting operations. Second, two different ciphering algorithms are used to encrypt the stego-image to obtain two different hybrid image security systems. The first encryption algorithm is based on binary code decomposition, while the second algorithm is a modification of an advanced encryption standard. After evaluating each hybrid algorithm alone, a comparison between the two hybrid systems is introduced to determine the best system. Several parameters were used for the performance, including the visual scene, histogram analysis, entropy, security analysis, and execution time.

     

Image quality assessment based on deep learning with FPGA implementation

To improve image quality assessment (IQA) methods, it is believable that we have to extract image features that are highly representative to human visual perception. In this paper, we propose a novel IQA algorithm by leveraging an optimized convolutional neural network architecture that is designed to automatically extract discriminative image quality features. And the IQA algorithm uses local luminance coefficient normalization, dropout and the other advanced techniques to further improve the network learning ability. At the same time the proposed IQA algorithm is implemented based on Field Programmable Gate Array (FPGA) and further evaluated on two public databases. Extensive experimental results have shown that our method outperforms many existing IQA algorithms in terms of accuracy and speed.     

A novel legitimacy preserving data hiding scheme based on LAS compressed code of VQ index tables

Data hiding is a technique for secret and secure data storing and transmission that embeds data into a media such as an image, audio, video and so on, with minimal quality degradation of the media. Some developed data hiding schemes are reversible. Reversibility property allows the media to be recovered completely after extraction of the embedded data. Vector Quantization (VQ)-based image data hiding is one of the most popular study areas in the literature. However, most VQ-based reversible data hiding schemes generate non-legitimate codes as output. In other words output codes generated by such schemes could not be decoded by the conventional VQ or VQ based decoders and may arouse the attention of interceptors. On the other hand, the existing VQ based reversible data hiding schemes that generate legitimate VQ codes as output, suffer from low capacity and poor quality of stego-image. In this paper a novel reversible data hiding scheme for VQ-compressed images based on locally adaptive data compression scheme (LAS) is proposed. Unlike other schemes, the proposed scheme doesn’t change the VQ indices; data is embedded by choosing one of the possible ways to encode each index. As a result, in comparison with the schemes that embed data by index replacement, in the proposed scheme no extra distortion is made by data embedding and the outputted codes are compatible with the conventional LAS decoder. These properties help to hide the existence of secret data and make the scheme suitable for steganography. Moreover, a framework to combine the proposed scheme with some other schemes to improve their capacity and embedding side information is proposed. Since LAS is a general data compression scheme, the proposed scheme could be used to embed data into any data formats. All existing LAS based data hiding schemes produce non-legitimate codes as their outputs and the proposed scheme is the first and only one that produces legitimate codes as output. Experimental results show that the proposed scheme outperforms the existing LAS based schemes and some other VQ based data hiding schemes. On average, the proposed scheme embeds 2.14 bits per index with almost the same bit-rate as the bit-rate of the VQ index table.     

基于小波分析的图像稀疏保真度评价

该文针对传统的图像质量评价方法无法有效模拟人类视觉系统(HVS)存在的不足,提出基于小波分析的加权稀疏保真度(Weighting Sparse Fidelity, WSF)图像评价算法。算法以模拟人类视觉系统的神经网络为切入点,对图像进行一阶小波分解得到4个不同方向的子带图像,然后将子带图像分成88大小的图像块,采用快速独立分量分析(FastICA)的方法对各个图像块进行训练并提取图像特征检测矩阵,根据特征检测矩阵计算各子带图像块的稀疏特征值并建立稀疏保真度质量评价模型。在此基础上,根据细节信息的不同对低频子带图像进行区间划分并设置视觉权重,使之更加接近人眼的主观视觉。实验中对LIVE库中所有图像进行算法验证,其结果表明,所提方法能很好地对各种失真类型的图像进行评价。基于小波分析的稀疏保真度评价算法能够有效模拟人类视觉系统的多频特性和视觉皮层感知机制,弥补现有图像质量评价方法在此方面的不足。