判别增强的生成对抗模型在文本至图像生成中的研究与应用

目前大部分基于生成对抗网络GAN的文本至图像生成算法着眼于设计不同模式的注意力生成模型,以提高图像细节的刻画与表达,但忽略了判别模型对局部关键语义的感知,以至于生成模型可能生成较差的图像细节“欺骗“判别模型。提出了判别语义增强的生成对抗网络DE-GAN模型,试图在判别模型中设计词汇-图像判别注意力模块,增强判别模型对关键语义的感知和捕捉能力,驱动生成模型生成高质量图像细节。实验结果显示,在CUB-Bird数据集上,DE-GAN在IS指标上达到了4.70,相比基准模型提升了4.2%,达到了较高的性能表现。     

Re-GAN：残差生成式对抗网络算法

目的 生成式对抗网络（generative adversarial network,GAN）是一种无监督生成模型,通过生成模型和判别模型的博弈学习生成图像。GAN的生成模型是逐级直接生成图像,下级网络无法得知上级网络学习的特征,以至于生成的图像多样性不够丰富。另外,随着网络层数的增加,参数变多,反向传播变得困难,出现训练不稳定和梯度消失等问题。针对上述问题,基于残差网络（residual network,ResNet）和组标准化（group normalization,GN）,提出了一种残差生成式对抗网络（residual generative adversarial networks,Re-GAN）。方法 Re-GAN在生成模型中构建深度残差网络模块,通过跳连接的方式融合上级网络学习的特征,增强生成图像的多样性和质量,改善反向传播过程,增强生成式对抗网络的训练稳定性,缓解梯度消失。随后采用组标准化（GN）来适应不同批次的学习,使训练过程更加稳定。结果 在Cifar10、CelebA和LSUN数据集上对算法的性能进行测试。Re-GAN的IS （inception score）均值在批次为64时,比DCGAN （deep convolutional GAN）和WGAN （Wasserstein-GAN）分别提高了5%和30%,在批次为4时,比DCGAN和WGAN分别提高了0.2%和13%,表明无论批次大小,Re-GAN生成图像具有很好的多样性。Re-GAN的FID （Fréchet inception distance）在批次为64时比DCGAN和WGAN分别降低了18%和11%,在批次为4时比DCGAN和WGAN分别降低了4%和10%,表明Re-GAN生成图像的质量更好。同时,Re-GAN缓解了训练过程中出现的训练不稳定和梯度消失等问题。结论 实验结果表明,在图像生成方面,Re-GAN的生成图像质量高、多样性强;在网络训练方面,Re-GAN在不同批次下的训练具有更好的兼容性,使训练过程更加稳定,梯度消失得到缓解。     

Data-driven generative design for mass customization: A case study

Generative design provides a promising algorithmic solution for mass customization of products, improving both product variety and design efficiency. However, the current designer-driven generative design formulates the automated program in a manual manner and has insufficient ability to satisfy the diverse needs of individuals. In this work, we propose a data-driven generative design framework by integrating multiple types of data to improve the automation level and performance of detail design to boost design efficiency and improve user satisfaction. A computational workflow including automated shape synthesis and structure design methods is established. More specifically, existing designs selected based on user preferences are utilized in the shape synthesis for creating generative models. For structural design, user-product interaction data gathered by sensors are used as inputs for controlling the spatial distributions of heterogeneous lattice structures. Finally, the proposed concept and workflow are demonstrated with a bike saddle design with a personalized shape and inner structures to be manufactured with additive manufacturing.     

A practical generative design method

A generative CAD based design exploration method is proposed. It is suitable for complex multi-criteria design problems where important performance criteria are uncomputable. The method is based on building a genotype of the design within a history based parametric CAD system and then, varying its parameters randomly within pre-defined limits to generate a set of distinctive designs. The generated designs are then filtered through various constraint envelopes representing geometric viability, manufacturability, cost and other performance related constraints, thus reducing the vast design space into a smaller viable design space represented by a set of distinctive designs. These designs may then be further developed by the designer. The proposed generative design method makes minimal imposition on the designer’s work process and maintains both flexibility and fluidity that is required for creative design exploration. Its ability to work seamlessly with current CAD based design practices from early conceptual to detailed design is demonstrated. The design philosophy behind this generative method and the key steps involved in its implementation are presented with examples.     

多模式生成对抗网络

生成对抗网络已经成为深度学习领域最热门的研究方向之一,其最大的优势在于能够以无监督的方式来拟合一个未知的分布。目前,生成对抗网络在图像生成领域大放异彩,其能够产生一些高质量的图像,但也暴露了一些弊端。在生成图像的过程中,经常会出现模式坍塌问题,从而导致生成的样本过于单一。为了解决这个问题,对生成对抗网络的模型结构和损失函数加以改进,使判别器能够从多个角度来度量生成数据的分布和真实数据的分布之间的差异,从而改善了生成样本的多样性。通过在多个数据集上进行实验,结果显示,提出的模型在很大程度上缓解了模式坍塌问题。     

Unpaired image to image transformation via informative coupled generative adversarial networks

We consider image transformation problems, and the objective is to translate images from a source domain to a target one. The problem is challenging since it is difficult to preserve the key properties of the source images, and to make the details of target being as distinguishable as possible. To solve this problem, we propose an informative coupled generative adversarial networks (ICoGAN). For each domain, an adversarial generator-and-discriminator network is constructed. Basically, we make an approximately-shared latent space assumption by a mutual information mechanism, which enables the algorithm to learn representations of both domains in unsupervised setting, and to transform the key properties of images from source to target.Moreover, to further enhance the performance, a weightsharing constraint between two subnetworks, and different level perceptual losses extracted from the intermediate layers of the networks are combined. With quantitative and visual results presented on the tasks of edge to photo transformation, face attribute transfer, and image inpainting, we demonstrate the ICo- GAN’s effectiveness, as compared with other state-of-the-art algorithms.     

Evolutionary Design Systems and Generative Processes

Design tools that aim not only to analyse and evaluate, but also to generate and explore alternative design proposals are now under development. An evolutionary paradigm is presented as a basis for creating such tools. First, the evolutionary paradigm is shown to be the only successful design system on which this new phase of design tool could be based. Secondly, any characterisation of design as a search problem is argued to be a serious misconception. Instead it is proposed that evolutionary design systems should be seen as generative processes that are able to evaluate their own output. Thirdly, a generic framework for generative evolutionary design systems is presented. Fourth, the generative process is introduced as a key element within this generic framework. The role of the environment within this process is fundamental. Finally, the direction of future research within the evolutionary design paradigm is discussed with possible short and long term goals being presented.     

条件生成对抗网络的低照度图像增强方法

目的 在日常的图像采集工作中,由于场景光照条件差或设备的补光能力不足,容易产生低照度图像。为了解决低照度图像视觉感受差、信噪比低和使用价值低（难以分辨图像内容）等问题,本文提出一种基于条件生成对抗网络的低照度图像增强方法。方法 本文设计一个具备编解码功能的卷积神经网络（CNN）模型作为生成模型,同时加入具备二分类功能的CNN作为判别模型,组成生成对抗网络。在模型训练的过程中,以真实的亮图像为条件,依靠判别模型监督生成模型以及结合判别模型与生成模型间的相互博弈,使得本文网络模型具备更好的低照度图像增强能力。在本文方法使用过程中,无需人工调节参数,图像输入模型后端到端处理并输出结果。结果 将本文方法与现有方法进行比较,利用本文方法增强的图像在亮度、清晰度以及颜色还原度等方面有了较大的提升。在峰值信噪比、直方图相似度和结构相似性等图像质量评价指标方面,本文方法比其他方法的最优值分别提高了0.7 dB、3.9%和8.2%。在处理时间上,本文方法处理图像的速度远远超过现有的传统方法,可达到实时增强的要求。结论 通过实验比较了本文方法与现有方法对于低照度图像的处理效果,表明本文方法具有更优的处理效果,同时具有更快的处理速度。     

The organizational fit of information systems

Abstract. One can expect that the fit, match or congruence between an information system and its organizational context will be of increasing importance as information systems become more and more integral parts of organizations. Existing IS research has largely neglected the theoretical analysis of the concept 'fit', however, even though it has been widely used in a 'pre-theoretical' sense. The analysis of the concept of 'organizational fit of an information system' put forward in this article is essentially based on the recent analysis of 'fit' in contingency theory, suggesting three interpretations for the concept: a selection approach, an interaction approach and a systems approach. A review of the IS literature on the organizational fit of information systems indicates a lack of any cumulative tradition in existing empirical contingency research into information systems and dominance of the selection approach. Because of the two limitations, this paper regards any definite assessment of the potential prospects of contingency research into information systems as overhasty and proposes a framework for future research into the situation dependence and organizational fit of information systems which integrates contingency research into information systems, contingency research into IS impacts and contingency research into the IS adoption (development/implementation) process.     

人脸年龄合成的生成对抗网络方法综述

年龄信息作为人类生物特征识别的重要组成部分,在社会保障和数字娱乐等领域具有广泛的应用前景。人脸年龄合成技术由于其广泛的应用价值,受到了越来越多学者的重视,已经成为计算机视觉领域的重要研究方向之一。随着深度学习的快速发展,基于生成对抗网络的人脸年龄合成技术已成为研究热点。尽管基于生成对抗网络的人脸年龄合成方法取得了不错的成果,但生成的人脸年龄图像仍存在图像质量较差、真实感较低、年龄转换效果和多样性不足等问题。主要因为当前人脸年龄合成研究仍存在以下困难： 1）现有人脸年龄合成数据集的限制; 2）引入人脸年龄合成的先验知识不足; 3）人脸年龄图像的细粒度性被忽视; 4）高分辨率下的人脸年龄合成问题;5）目前人脸年龄合成方法的评价标准不规范。本文对目前人脸年龄合成技术进行全面综述,以人脸年龄合成方法为研究对象,阐述其研究现状。通过调研文献,对人脸年龄合成方法进行分类,重点介绍了基于生成对抗网络的人脸年龄合成方法。此外,本文还讨论了常用的人脸年龄合成数据集及评价指标,分析了各种人脸年龄合成方法的基本思想、特点及其局限性,对比了部分代表方法的性能,指出了该领域目前存在的挑战并提供了一些具有潜力的研究方向,为研究者们解决存在的问题提供便利。     

Generation of energy-efficient architecture solutions applying GENE_ARCH: An evolution-based generative design system

As the field of design automation and generative design systems (GDS) evolve, more emphasis is placed on issues of design evaluation. This paper focus on the presentation of different applications of GENE_ARCH, an evolution-based GDS aimed at helping architects to achieve energy-efficient and sustainable architectural solutions. The system applies goal-oriented design, combining a genetic algorithm (GA) as the search engine, with the DOE2.1E building energy simulation software as the evaluation module. Design evaluation is based on energy spent for heating, cooling, ventilation and artificial lighting in the building, and on sustainability issues like greenhouse gas emissions associated with the embodied energy of construction materials. The GA can work either as a standard GA or as a Pareto GA, for multicriteria search and optimization. In order to provide a broad view of the capabilities of the software, different applications are discussed: (1) standard GA: testing and validating the software; (2) standard GA: incorporation of architecture design intentions, using a building by architect Alvaro Siza; (3) Pareto GA: choice of construction materials, considering cost, building energy use, and embodied energy; (4) Pareto GA: application to Siza’s building, considering thermal and lighting behavior separately; (5) standard GA: shape generation with single objective function; (6) Pareto GA: shape generation with multicriteria evaluation; (7) Pareto GA: application to an urban and housing context. Overall conclusions from the different applications are discussed, as well as current challenges and limitations, and directions for further work.     

基于生成艺术的空间形态生成研究

有着2500年历史的绍兴市,以纵横交错的河道、众多的石桥和乌篷船而闻名于世。本文基于生成艺术的思想,采用GIS技术和三维建模技术生成了一种绍兴市独有的空间形态（乌篷形态）,该形态对于计算机创作和建筑设计领域具有积极的启发意义。     

生成对抗网络研究综述

首先介绍了生成对抗网络基本理论、应用场景和研究现状,并列举了其亟待改进的问题.围绕针对提升模型训练效率、提升生成样本质量和降低模式崩溃现象发生可能性3类问题的解决,从模型结构和训练过程两大改进方向和7个细分维度,对近年来生成对抗网络的主要研究工作、改进机理和特点进行了归纳和总结,并结合3方面对其未来的研究方向进行了探讨...     

基于多谱特征生成对抗网络的语音转换算法

语音转换在教育、娱乐、医疗等各个领域都有广泛的应用,为了得到高质量的转换语音,提出了基于多谱特征生成对抗网络的语音转换算法。利用生成对抗网络对由谱特征参数生成的声纹图进行转换,利用特征级多模态融合技术使网络学习来自不同特征域的多种信息,以提高网络对语音信号的感知能力,从而得到具有良好清晰度和可懂度的高质量转换语音。实验结果表明,在主、客观评价指标上,本文算法较传统算法均有明显提升。     

融合参考先验与生成先验的老照片修复

目的 修复老照片具有重要的实用价值,但老照片包含多种未知复杂的退化,传统的修复方法组合不同的数字图像处理技术进行修复,通常产生不连贯或不自然的修复结果。基于深度学习的修复方法虽然被提出,但大多集中于对单一或有限的退化进行修复。针对上述问题,本文提出一种融合参考先验与生成先验的生成对抗网络来修复老照片。方法 对提取的老照片和参考图像的浅层特征进行编码获得深层语义特征与潜在编码,对获得的潜在编码进一步融合获得深度语义编码,深度语义编码通过生成先验网络获得生成先验特征,并且深度语义编码引导条件空间多特征变换条件注意力块进行参考语义特征、生成先验特征与待修复特征的空间融合变换,最后通过解码网络重建修复图像。结果 实验与6种图像修复方法进行了定量与定性评估。比较了4种评估指标,本文算法的所有指标评估结果均优于其他算法,PSNR (peak signal-to-noise ratio)为23.69 d B,SSIM (structural similarity index)为0.828 3,FID (Fréchet inception distance)为71.53,LPIPS (learned ...     

Cancer classification with data augmentation based on generative adversarial networks

Accurate diagnosis is a significant step in cancer treatment. Machine learning can support doctors in prognosis decision-making, and its performance is always weakened by the high dimension and small quantity of genetic data. Fortunately, deep learning can effectively process the high dimensional data with growing. However, the problem of inadequate data remains unsolved and has lowered the performance of deep learning. To end it, we propose a generative adversarial model that uses non target cancer data to help target generator training. We use the reconstruction loss to further stabilize model training and improve the quality of generated samples. We also present a cancer classification model to optimize classification performance. Experimental results prove that mean absolute error of cancer gene made by our model is 19.3% lower than DC-GAN, and the classification accuracy rate of our produced data is higher than the data created by GAN. As for the classification model, the classification accuracy of our model reaches 92.6%, which is 7.6% higher than the model without any generated data.     

A generative design technique for exploring shape variations

Because innovative and creative design is essential to a successful product, this work brings the benefits of generative design in the conceptual phase of the product development process so that designers/engineers can effectively explore and create ingenious designs and make better design decisions. We proposed a state-of-the-art generative design technique (GDT), called Space-filling-GDT (Sf-GDT), for the creation of innovative designs. The proposed Sf-GDT has the ability to create variant optimal design alternatives for a given computer-aided design (CAD) model. An effective GDT should generate design alternatives that cover the entire design space. Toward that end, the criterion of space-filling is utilized, which uniformly distribute designs in the design space thereby giving a designer a better understanding of possible design options. To avoid creating similar designs, a weighted-grid-search approach is developed and integrated into the Sf-GDT. One of the core contributions of this work lies in the ability of Sf-GDT to explore hybrid design spaces consisting of both continuous and discrete parameters either with or without geometric constraints. A parameter-free optimization technique, called Jaya algorithm, is integrated into the Sf-GDT to generate optimal designs. Three different design parameterization and space formulation strategies; explicit, interactive, and autonomous, are proposed to set up a promising search region(s) for optimization. Two user interfaces; a web-based and a Windows-based, are also developed to utilize Sf-GDT with the existing CAD software having parametric design abilities. Based on the experiments in this study, Sf-GDT can generate creative design alternatives for a given model and outperforms existing state-of-the-art techniques.     

Towards physical interaction-based sequential mobility assistance using latent generative model of movement state

ABSTRACT

In this paper, we address sequential mobility assistance for daily elderly care through physical human–robot interaction. The goal of this work is to develop a robotic assistive system to provide physical support in daily life such as movement transition, e.g. sit-to-stand and walking. Using a mobile human support robotic platform, we propose an unsupervised learning-based approach to providing desirable physical support through an adaptive impedance parameter selection strategy according to the recognized user's movement state in an online manner. Using a latent generative model with a long short-term memory-based variational autoencoder, we first estimate the probability of the user's current movement state based on the sensory information in a low dimensional latent space. Then, the desired impedance parameters are selected adaptively according to the estimated movement state. One of the benefits of such an unsupervised learning approach is that no labeling is necessary in the training phase. Furthermore, our proposed framework is capable of detecting possible novel states such as falling over based on the obtained latent space. In order to demonstrate the proof of concept of our proposed approach, we present the experimental results of performance evaluations of online movement state recognition as well as novel movement detection.     

联合生成与判别模型的目标检测与跟踪

提出了一种新的基于生成-判别模型的目标检测与跟踪方法。利用DAISY特征描述子所具有的对光照、形变、视角、尺度的不变性以及计算高效的特点,提取目标稳定的特征点并加以表达,形成生成模型;采用霍夫森林分类器作为判别模型,用以训练目标图像块。在后续视频序列中利用目标的检测结果和判别码本的相似性测量对模型进行更新,构建一个动态自适应的判别码本。实验证明这种将快速有效的DAISY描述子和识别率高、鲁棒性强的霍夫森林分类器相结合的算法,跟踪精度高、实时性较好,具有目标局部防遮挡能力和不同分辨率下的识别能力。