Generative Adversarial Networks:Introduction and Outlook

Recently, generative adversarial networks(GANs)have become a research focus of artificial intelligence. Inspired by two-player zero-sum game, GANs comprise a generator and a discriminator, both trained under the adversarial learning idea.The goal of GANs is to estimate the potential distribution of real data samples and generate new samples from that distribution.Since their initiation, GANs have been widely studied due to their enormous prospect for applications, including image and vision computing, speech and language processing, etc. In this review paper, we summarize the state of the art of GANs and look into the future. Firstly, we survey GANs' proposal background,theoretic and implementation models, and application fields.Then, we discuss GANs' advantages and disadvantages, and their development trends. In particular, we investigate the relation between GANs and parallel intelligence,with the conclusion that GANs have a great potential in parallel systems research in terms of virtual-real interaction and integration. Clearly, GANs can provide substantial algorithmic support for parallel intelligence.

Multi-distortion suppression for neutron radiographic images based on generative adversarial network

Neutron radiography is a crucial nondestructive testing technology widely used in the aerospace,military,and nuclear industries.However,because of the physical limitations of neutron sources and collimators,the resulting neutron radiographic images inevitably exhibit multiple distortions,including noise,geometric unsharpness,and white spots.Furthermore,these distortions are particularly significant in compact neutron radiography systems with low neutron fluxes.Therefore,in this study,we devised a multi-distortion suppression network that employs a modified generative adversarial network to improve the quality of degraded neutron radiographic images.Real neutron radiographic image datasets with various types and levels of distortion were built for the first time as multi-distortion suppression datasets.Thereafter,the coordinate attention mechanism was incorporated into the backbone network to augment the capability of the proposed network to learn the abstract relationship between ideally clear and degraded images.Extensive experiments were performed;the results show that the proposed method can effectively suppress multiple distortions in real neutron radiographic images and achieve state-of-theart perceptual visual quality,thus demonstrating its application potential in neutron radiography.

General image classification method based on semi-supervised generative adversarial networks

Generative adversarial networks(GANs) have become a competitive method among computer vision tasks. There have been many studies devoted to utilizing generative network to do generative tasks, such as images synthesis. In this paper, a semi-supervised learning scheme is incorporated with generative adversarial network on image classification tasks to improve the image classification accuracy. Two applications of GANs are mainly focused on: semi-supervised learning and generation of images which can be as real as possible. The whole process is divided into two sections. First, only a small part of the dataset is utilized as labeled training data. And then a huge amount of samples generated from the generator is added into the training samples to improve the generalization of the discriminator. Through the semi-supervised learning scheme, full use of the unlabeled data is made which may contain potential information. Thus, the classification accuracy of the discriminator can be improved. Experimental results demonstrate the improvement of the classification accuracy of discriminator among different datasets, such as MNIST, CIFAR-10.

Ballistic response of armour plates using Generative Adversarial Networks

It is important to understand how ballistic materials respond to impact from projectiles such that informed decisions can be made in the design process of protective armour systems. Ballistic testing is a standards-based process where materials are tested to determine whether they meet protection, safety and performance criteria. For the V50ballistic test, projectiles are fired at different velocities to determine a key design parameter known as the ballistic limit velocity(BLV), the velocity above which projectiles perforate the target. These tests, however, are destructive by nature and as such there can be considerable associated costs, especially when studying complex armour materials and systems. This study proposes a unique solution to the problem using a recent class of machine learning system known as the Generative Adversarial Network(GAN). The GAN can be used to generate new ballistic samples as opposed to performing additional destructive experiments. A GAN network architecture is tested and trained on three different ballistic data sets, and their performance is compared. The trained networks were able to successfully produce ballistic curves with an overall RMSE of between 10 and 20 % and predicted the V50BLV in each case with an error of less than 5 %. The results demonstrate that it is possible to train generative networks on a limited number of ballistic samples and use the trained network to generate many new samples representative of the data that it was trained on. The paper spotlights the benefits that generative networks can bring to ballistic applications and provides an alternative to expensive testing during the early stages of the design process.

Evolution and Effectiveness of Loss Functions in Generative Adversarial Networks

Recently,the evolution of Generative Adversarial Networks(GANs)has embarked on a journey of revolutionizing the field of artificial and computational intelligence.To improve the generating ability of GANs,various loss functions are introduced to measure the degree of similarity between the samples generated by the generator and the real data samples,and the effectiveness of the loss functions in improving the generating ability of GANs.In this paper,we present a detailed survey for the loss functions used in GANs,and provide a critical analysis on the pros and cons of these loss functions.First,the basic theory of GANs along with the training mechanism are introduced.Then,the most commonly used loss functions in GANs are introduced and analyzed.Third,the experimental analyses and comparison of these loss functions are presented in different GAN architectures.Finally,several suggestions on choosing suitable loss functions for image synthesis tasks are given.

Exploration of the Relation between Input Noise and Generated Image in Generative Adversarial Networks

In this paper,we propose a hybrid model aiming to map the input noise vector to the label of the generated image by the generative adversarial network(GAN).This model mainly consists of a pre-trained deep convolution generative adversarial network(DCGAN)and a classifier.By using the model,we visualize the distribution of two-dimensional input noise,leading to a specific type of the generated image after each training epoch of GAN.The visualization reveals the distribution feature of the input noise vector and the performance of the generator.With this feature,we try to build a guided generator(GG)with the ability to produce a fake image we need.Two methods are proposed to build GG.One is the most significant noise(MSN)method,and the other utilizes labeled noise.The MSN method can generate images precisely but with less variations.In contrast,the labeled noise method has more variations but is slightly less stable.Finally,we propose a criterion to measure the performance of the generator,which can be used as a loss function to effectively train the network.

融合IMR-WGAN的时序数据修复方法

工业数据由于技术故障和人为因素通常导致数据异常,现有基于约束的方法因约束阈值设置的过于宽松或严格会导致修复错误,基于统计的方法因平滑修复机制导致对时间步长较远的异常值修复准确度较低.针对上述问题,提出了基于奖励机制的最小迭代修复和改进WGAN混合模型的时序数据修复方法.首先,在预处理阶段,保留异常数据,进行信息标注等处理,从而充分挖掘异常值与真实值之间的特征约束.其次,在噪声模块提出了近邻参数裁剪规则,用于修正最小迭代修复公式生成的噪声向量.将其传递至模拟分布模块的生成器中,同时设计了一个动态时间注意力网络层,用于提取时序特征权重并与门控循环单元串联组合捕捉不同步长的特征依赖,并引入递归多步预测原理共同提升模型的表达能力;在判别器中设计了Abnormal and Truth奖励机制和Weighted Mean Square Error损失函数共同反向优化生成器修复数据的细节和质量.最后,在公开数据集和真实数据集上的实验结果表明,该方法的修复准确度与模型稳定性显著优于现有方法.

MACDCGAN的发电机轴承故障诊断方法

在实际工况中,发电机中传感器采集到的故障样本数据有限,使用基于深度学习的方法进行故障诊断存在过拟合问题导致模型泛化能力较差以及诊断精度不高。为了解决这个问题,采用样本扩充的思路,提出了一种改进的辅助分类器条件深度卷积生成对抗网络(MACDCGAN)的故障诊断方法。通过对采集的一维时序信号进行小波变换增强特征,构建简化结构参数的条件深度卷积生成对抗网络模型生成样本,并在模型中采用Wasserstein距离优化损失函数解决训练过程中存在模式崩塌和梯度消失的缺点;通过添加一个独立的分类器来改进分类模型的兼容性,并在分类器中引入学习率衰减算法增加模型稳定性。试验结果表明,该方法可以有效地提高故障诊断的精度,并且验证了所提模型具有良好的泛化性能。

WGAN-GP结合CBAM-VGG16轻量化网络滚动轴承故障诊断

针对滚动轴承故障诊断中存在的故障数据不平衡且诊断效率低的问题,提出一种将改进的生成对抗网络(Wasserstein Generative Adversarial Network with Gradient Penalty,WGAN-GP)和轻量化卷积神经网络相结合的滚动轴承故障诊断方法。首先,利用连续小波变换(Continuous Wavelet Transform,CWT)生成二维时频图,并通过WGAN-GP进行数据增强;其次在视觉几何群网络(Visual Geometry Group Network-16,VGG16)的基础上,引入Ghost模块和全局平均池化(Global Average Pooling,GAP)对其进行轻量化改进;再次,利用卷积注意力模块(Convolutional Block Attention Module,CBAM)和带重启的余弦退火衰减法提高VGG16模型的性能,构建CBAM-VGG16轻量化卷积神经网络模型,将增强后的数据进行预处理后输入到模型中进行训练,建立故障诊断模型;最后采用西储大学轴承数据集进行模型验证和分析。实验结果表明:该方法证实了故障数据不足时进行故障诊断的可行性,缩短了模型的训练时间、诊断时间并缩减了模型的大小和参数量,提高了故障诊断的效率和准确率。

双分支GAN与注意力机制的火灾隐患检测算法

针对传统火灾报警在夜间等极端天气下效果不佳,受限于复杂环境等问题,提出一种基于红外与可见光图像融合的火灾预警算法。在生成对抗网络(GAN)中设计并提出双分支注意力结构。其中一条分支通过密集残差子网提取更多鲁棒的特征信息,另一条分支通过注意力子网(efficient coordinate channel attention group,ECCAG)弥补空间信息的缺失,以最大限度获取更多高频细节特征,设计并提出了一种调节损失作为损失函数,通过改进GAN算法得到融合图像,根据提出的火灾预警算法判断是否存在火灾隐患。实验结果表明:改进GAN算法得到的融合数据集目标检测的平均准确率为96.19%,相较于单一红外数据集与原始GAN算法数据集的目标检测平均准确率分别提高了11.09个百分点与6.2个百分点,在公开数据集TNO与LLVIP数据集上测试火灾患检测准确率为97.45%。结果表明,火灾预警算法可以在未发生火灾时及时预警,针对不同场景都可得到显著的检测效果。

基于Transformer和GAN的对抗样本生成算法

对抗攻击与防御是计算机安全领域的一个热门研究方向。针对现有基于梯度的对抗样本生成方法可视质量差、基于优化的方法生成效率低的问题,提出基于Transformer和生成对抗网络(GAN)的对抗样本生成算法Trans-GAN。首先利用Transformer强大的视觉表征能力,将其作为重构网络,用于接收干净图像并生成攻击噪声;其次将Transformer重构网络作为生成器,与基于深度卷积网络的鉴别器相结合组成GAN网络架构,提高生成图像的真实性并保证训练的稳定性,同时提出改进的注意力机制Targeted Self-Attention,在训练网络时引入目标标签作为先验知识,指导网络模型学习生成具有特定攻击目标的对抗扰动;最后利用跳转连接将对抗噪声施加在干净样本上,形成对抗样本,攻击目标分类网络。实验结果表明:Trans-GAN算法针对MNIST数据集中2种模型的攻击成功率都达到99.9%以上,针对CIFAR10数据集中2种模型的攻击成功率分别达到96.36%和98.47%,优于目前先进的基于生成式的对抗样本生成方法;相比快速梯度符号法和投影梯度下降法,Trans-GAN算法生成的对抗噪声扰动量更小,形成的对抗样本更加自然,满足人类视觉不易分辨的要求。

Generative Adversarial Network Based Heuristics for Sampling-Based Path Planning

Sampling-based path planning is a popular methodology for robot path planning.With a uniform sampling strategy to explore the state space,a feasible path can be found without the complex geometric modeling of the configuration space.However,the quality of the initial solution is not guaranteed,and the convergence speed to the optimal solution is slow.In this paper,we present a novel image-based path planning algorithm to overcome these limitations.Specifically,a generative adversarial network(GAN)is designed to take the environment map(denoted as RGB image)as the input without other preprocessing works.The output is also an RGB image where the promising region(where a feasible path probably exists)is segmented.This promising region is utilized as a heuristic to achieve non-uniform sampling for the path planner.We conduct a number of simulation experiments to validate the effectiveness of the proposed method,and the results demonstrate that our method performs much better in terms of the quality of the initial solution and the convergence speed to the optimal solution.Furthermore,apart from the environments similar to the training set,our method also works well on the environments which are very different from the training set.

基于CSLS-CycleGAN的侧扫声纳水下目标图像样本扩增法

针对侧扫声纳水下目标图像稀缺,获取难度大、成本高,导致基于深度学习的目标检测模型性能差的问题,结合光学域类目标数据集丰富的现状,提出一种基于通道和空间注意力(channel and spatial attention,CSA)模块、最小二乘生成对抗生成网络(least squares generative adversarial networks,LSGAN)及循环对抗生成网络(cycle generative adversarial networks,CycleGAN)的侧扫声纳水下目标图像样本扩增方法。首先,受CycleGAN的启发,设计基于循环一致性的单循环网络结构,保证模型的训练效率。然后,在生成器中融合CSA模块,减少信息弥散的同时增强跨纬度交互。最后,设计了基于LSGAN的损失函数,提高生成图像质量的同时提高训练稳定性。在船舶光学域数据集与侧扫声纳沉船数据集上进行实验,所提方法实现了光学-侧扫声纳样本间信息的高效、稳健转换以及大量侧扫声纳目标样本的扩增。同时,基于本文生成样本训练后的检测模型进行了水下目标检测,结果表明,使用本文样本扩增数据训练后的模型在少样本沉船目标检测的平均准确率达到了84.71%,证明了所提方法实现了零样本和小样本水下强代表性目标样本的高质量扩增,并为高性能水下目标检测模型构建提供了一种新的途径。

基于有效注意力和GAN结合的脑卒中EEG增强算法

在基于脑电的卒中分类诊断任务中,以卷积神经网络为基础的深度模型得到广泛应用,但由于卒中类别病患样本数量少,导致数据集类别不平衡,降低了分类精度。现有的少数类数据增强方法大多采用生成对抗网络(GAN),生成效果一般,虽然可通过引入缩放点乘注意力改善样本生成质量,但存储及运算代价往往较大。针对此问题,构建一种基于线性有效注意力的渐进式数据增强算法LESA-CGAN。首先,算法采用双层自编码条件生成对抗网络架构,分别进行脑电标签特征提取及脑电样本生成,并使生成过程逐层精细化;其次,通过在编码部分引入线性有效自注意力(LESA)模块,加强脑电的标签隐层特征提取,并降低网络整体的运算复杂度。消融与对比实验结果表明,在合理的编码层数与生成数据比例下,LESA-CGAN与其他基准方法相比计算资源占用较少,且在样本生成质量指标上实现了10%的性能提升,各频段生成的脑电特征样本均更加自然,同时将病患分类的准确率和敏感度提高到了98.85%和98.79%。

融合门控变换机制和GAN的低光照图像增强方法

针对低光照图像增强过程中存在的配对图像数据依赖、细节损失严重和噪声放大问题,提出结合门控通道变换机制和生成对抗网络(GAN)的低光照图像增强方法AGR-GAN,该方法可以在没有低/正常光图像对的情况下进行训练。首先,设计特征提取网络,该网络由多个基于门控通道变换单元的多尺度卷积残差模块构成,以提取输入图像的全局上下文特征和多尺度局部特征信息;然后,在特征融合网络中,采用卷积残差结构将提取的深浅层特征进行充分融合,再引入横向跳跃连接结构,最大程度保留细节特征信息,获得最终的增强图像;最后,引入联合损失函数指导网络训练过程,抑制图像噪声,使增强图像色彩更自然匀称。实验结果表明,该方法在主观视觉分析和客观指标评价方面相较其他算法均具有显著优势,其能有效提高低光照图像的亮度和对比度,减弱图像噪声,增强后的图像更清晰且色彩更真实,峰值信噪比、结构相似度和无参考图像质量评价指标平均可达16.48 dB、0.93和3.37。

基于SASGAN的戏剧脸谱多样化生成

为解决现有自动生成的戏剧脸谱在分辨率和真实性上效果不佳的问题,提出了基于自注意力机制的风格化生成对抗网络(SASGAN)。首先在StyleGAN的基础上引入了自注意力机制以及矢量量化方法,增强了对脸谱图案几何结构特征的提取,接着通过多样化差异性增强(DDG)扩充数据,采用脸谱色调辅助算法对DDG方法进行补充,建立了包含12599张图像的戏剧脸谱数据集,最后在此数据集上进行训练,生成了兼顾多样性和真实性的脸谱图像。实验结果表明,对于戏剧脸谱图像,DDG方法较传统方法在数据增广方面有着较大提升,而SASGAN则提升了戏剧脸谱图像的分辨率和真实性,在主观视觉上得到了理想的效果。

Single Image Dehazing: An Analysis on Generative Adversarial Network

Haze is a very common phenomenon that degrades or reduces visibility. It causes various problems where high-quality images are required such as traffic and security monitoring. So haze removal from scenes is an immediate demand for clear vision. Recently, in addition to the conventional dehazing mechanisms, different types of deep generative adversarial networks (GAN) are applied to suppress the noise and improve the dehazing performance. But it is unclear how these algorithms would perform on hazy images acquired “in the wild” and how we could gauge the progress in the field. To bridge this gap, this presents a comprehensive study on three single image dehazing state-of-the-art GAN models, such as AOD-Net, cGAN, and DHSGAN. We have experimented using benchmark dataset consisting of both synthetic and real-world hazy images. The obtained results are evaluated both quantitatively and qualitatively. Among these techniques, the DHSGAN gives the best performance.

Solar image deconvolution by generative adversarial network

With aperture synthesis(AS)technique,a number of small antennas can be assembled to form a large telescope whose spatial resolution is determined by the distance of two farthest antennas instead of the diameter of a single-dish antenna.In contrast from a direct imaging system,an AS telescope captures the Fourier coefficients of a spatial object,and then implement inverse Fourier transform to reconstruct the spatial image.Due to the limited number of antennas,the Fourier coefficients are extremely sparse in practice,resulting in a very blurry image.To remove/reduce blur,“CLEAN”deconvolution has been widely used in the literature.However,it was initially designed for a point source.For an extended source,like the Sun,its efficiency is unsatisfactory.In this study,a deep neural network,referring to Generative Adversarial Network(GAN),is proposed for solar image deconvolution.The experimental results demonstrate that the proposed model is markedly better than traditional CLEAN on solar images.The main purpose of this work is visual inspection instead of quantitative scientific computation.We believe that this will also help scientists to better understand solar phenomena with high quality images.

Point cloud upsampling generative adversarial network based on residual multi-scale off-set attention

Background Owing to the limitations of the working principle of three-dimensional(3D) scanning equipment, the point clouds obtained by 3D scanning are usually sparse and unevenly distributed. Method In this paper, we propose a new generative adversarial network(GAN) that extends PU-GAN for upsampling of point clouds. Its core architecture aims to replace the traditional self-attention(SA) module with an implicit Laplacian offset attention(OA) module and to aggregate the adjacency features using a multiscale offset attention(MSOA)module, which adaptively adjusts the receptive field to learn various structural features. Finally, residual links are added to create our residual multiscale offset attention(RMSOA) module, which utilizes multiscale structural relationships to generate finer details. Result The results of several experiments show that our method outperforms existing methods and is highly robust.

基于改进MMD-GAN的可再生能源随机场景生成

针对可再生能源出力不确定性的准确表征问题,提出了一种基于改进的最大均值差异生成对抗网络(maximum mean discrepancy generative adversarial networks,MMD-GAN)的可再生能源随机场景生成方法。首先,阐述了GAN及MMD-GAN的基本原理,提出了MMD-GAN的改进方案,即在MMD-GAN的基础上改进鉴别器损失函数,并采用谱归一化和有界高斯核提升生成器和鉴别器的训练稳定性。然后,设计了基于改进MMD-GAN的可再生能源随机场景生成流程。最后,分析了所提方法在可再生能源随机场景生成中的效果,比较了改进MMD-GAN方法与MMD-GAN方法及典型GAN方法的性能差异。结果表明,改进MMD-GAN方法在生成分布和真实分布的Wasserstein距离上较对比方法降低超过50%,生成的场景精度得到有效提升。