Generalized zero-shot classification via iteratively generating and selecting unseen samples

Generalized zero-shot classification (GZSC) is a challenging task to recognize seen and unseen samples from target domain by seen samples in source domain. Since the lack of unseen data, many methods train a generative adversarial network (GAN) to generate unseen samples. However, the GAN model trained by seen samples is not suitable for generating unseen samples. For dealing with this problem, we train the GAN model by generating seen and unseen samples, simultaneously. In order to generate high-quality unseen samples, the visual prototypes of the generated unseen samples are made near to the real unseen visual prototypes. We select the confident unseen samples based on the agreement of the current two unseen classifiers and use them to update the unseen visual prototypes. Through the iteratively generating and selecting method (IGS), we can generate high-quality unseen samples and select the most confident unseen samples. Experimental results on the standard benchmarks show the superiority of the proposed model over the state-of-the-art methods for GZSC tasks.     

Semantic-aware visual attributes learning for zero-shot recognition

Zero-shot learning (ZSL) aims to recognize unseen image classes without requiring any training samples of these specific classes. The ZSL problem is typically achieved by building up a semantic embedding space like attributes to bridge the visual features and class labels of images. Currently, most ZSL approaches focus on learning a visual-semantic alignment from seen classes using only the human-designed attributes, and then ZSL problem is solved by transferring semantic knowledge from seen classes to the unseen classes. However, few works indicate if the human-designed attributes are discriminative enough for image class prediction. To address this issue, we propose a semantic-aware dictionary learning (SADL) framework to explore these discriminative visual attributes across seen and unseen classes. Furthermore, the semantic cues are elegantly integrated into the feature representations via learned visual attributes for recognition task. Experiments conducted on two challenging benchmark datasets show that our approach outweighs other state-of-the-art ZSL methods.     

基于半监督学习生成对抗网络的人脸还原算法研究

基于大量训练样本生成高置信度图像的生成对抗网络研究已经取得一些成果,但是现有的研究只针对已知训练样本进行图像生成,而未将训练的参数用于训练样本之外的图像生成。该文设计了一种改进的生成对抗网络模型,在已有网络的基础上增加一个还原层,使得测试图像可以通过改进的对抗网络生成对应的高置信度图像。实验结果表明,改进的生成对抗网络参数可以应用到训练集之外的普通样本。同时本文改进了生成模型的损失算法,极大地缩短了网络的收敛时间。     

基于空间语义对象混合学习的复杂图像场景自动分类方法研究

场景分类是将多幅图像标记为不同语义类别的过程。该文针对现有方法对复杂图像场景分类性能欠佳的不足,提出一种新的基于空间语义对象混合学习的复杂图像场景分类方法。该方法以多尺度分割得到的图像对象而非整幅图像为主体进行产生式语义建模,统计各类有效特征挖掘对象的类别分布信息,并通过空间金字塔匹配,构建包含层次数据和语义信息的中间向量,弥补语义鸿沟的缺陷,训练中还结合判别式学习提高分类器的可信性。在实验数据集上的结果表明该方法具备较高的学习性能和分类精度,适用于多种类型和复杂内容图像的解译,具有较强的实用价值。     

基于改进生成对抗网络和MobileNetV3的带钢缺陷分类

针对数据集样本数量较少会影响深度学习检测效果的问题,提出了一种基于改进生成对抗网络和MobileNetV3的带钢缺陷分类方法。首先,引入生成对抗网络并对生成器和判别器进行改进,解决了类别错乱问题并实现了带钢缺陷数据集的扩充。然后,对轻量级图像分类网络MobileNetV3进行改进。最后,在扩充后的数据集上训练,实现了带钢缺陷的分类。实验结果表明,改进的生成对抗网络可生成比较真实的带钢缺陷图像,同时解决深度学习中样本不足的问题;且改进的MobileNetV3参数量是原有参数量的1/14左右,准确率为94.67%,比改进前提高了2.62个百分点,可在工业现场对带钢缺陷进行实时准确的分类。     

基于生成对抗网络的舰船辐射噪声分类方法研究

基于机器学习的舰船目标识别近年来已成为水声信号处理领域的一个重要研究方向,但水声目标信号的获取困难,样本量不足和不均衡的问题很容易导致目标分类模型的识别效果不佳。该文提出一种基于条件卷积生成对抗网络的船舶噪声数据分类方法,该方法利用生成对抗学习理论,生成相比于传统数据增强算法非线性特征更强,特征差异更丰富的伪DEMON调制谱数据来缓解训练样本量不足的问题。之后将传统生成对抗网络中的全连层输出替换成更善于解决小样本问题集成分类器,从而降低分类器对于数据量的依赖程度,进一步提高分类模型性能。最终由基于真实样本的实验结果表明,相比于传统数据增强算法和卷积生成对抗网络,该文方法能够更有效提高在样本不足条件下的模型的分类性能。     

基于元学习的无监督风格迁移算法研究

最近,在生成式对抗网络和足够的非配对训练数据下,无监督领域风格迁移取得了较高的性能。然而，现有的领域迁移框架大多基于庞大的训练数据集，且只能根据训练图像进行特定类别的风格迁移，忽略了其中的学习经验被，使获得的模型不能适应新的领域。文中对传统的非配对循环生成对抗网络Cycle-GAN进行了改进，并使用元学习方法训练了无监督领域的风格迁移问题。另外，文中提出的模型在7个不同的双域迁移任务上证明了其有效性，当对每个新领域进行小样本训练时，该算法均优于传统的风格迁移算法。     

Semantic guided knowledge graph for large-scale zero-shot learning

Zero-shot learning has received growing attention, which aims to improve generalization to unseen concepts. The key challenge in zero-shot tasks is to precisely model the relationship between seen and unseen classes. Most existing zero-shot learning methods capture inter-class relationships via a shared embedding space, leading to inadequate use of relationships and poor performance. Recently, knowledge graph-based methods have emerged as a new trend of zero-shot learning. These methods use a knowledge graph to accurately model the inter-class relationships. However, the currently dominant method for zero-shot learning directly extracts the fixed connection from off-the-shelf WordNet, which will inherit the inherent noise in WordNet. In this paper, we propose a novel method that adopts class-level semantic information as a guidance to construct a new semantic guided knowledge graph (SG-KG), which can correct the errors in the existing knowledge graph and accurately model the inter-class relationships. Specifically, our method includes two main steps: noise suppression and semantic enhancement. Noise suppression is used to eliminate noise edges in the knowledge graph, and semantic enhancement is used to connect two classes with strong relations. To promote high efficient information propagation among classes, we develop a novel multi-granularity fusion network (MGFN) that integrates discriminative information from multiple GCN branches. Extensive experiments on the large-scale ImageNet-21K dataset and AWA2 dataset demonstrate that our method consistently surpasses existing methods and achieves a new state-of-the-art result.     

EdgeGAN: One-way mapping generative adversarial network based on the edge information for unpaired training set

Image conversion has attracted mounting attention due to its practical applications. This paper proposes a lightweight network structure that can implement unpaired training sets to complete one-way image mapping, based on the generative adversarial network (GAN) and a fixed-parameter edge detection convolution kernel. Compared with the cycle consistent adversarial network (CycleGAN), the proposed network features simpler structure, fewer parameters (only 37.48% of the parameters in CycleGAN), and less training cost (only 35.47% of the GPU memory usage and 17.67% of the single iteration time in CycleGAN). Remarkably, the cyclic consistency becomes not mandatory for ensuring the consistency of the content before and after image mapping. This network has achieved significant processing effects in some image translation tasks, and its effectiveness and validity have been well demonstrated through typical experiments. In the quantitative classification evaluation based on VGG-16, the algorithm proposed in this paper has achieved superior performance.     

生成对抗网络的发展与挑战

生成对抗网络（Generative adversarial network, GAN）由生成模型和判别模型构成,生成模型获取真实数据的概率分布,判别模型判断输入是真实数据还是生成器生成的数据,二者通过相互对抗训练,最终使生成模型学习到真实数据的分布,使判别模型无法准确判断输入数据的来源。生成对抗网络为视觉分类任务的算法性能的提升开辟了新的思路,自诞生之日起至今已经在各个领域产生了大量变体。本文的主要内容包括：生成对抗网络的研究现状、应用场景和基本模型架构,并列举了生成对抗网络本身所存在的弊端;从网络架构、损失函数和训练方式这三方面对生成对抗网络的各种主要典型发展进行归纳;详细总结和分析了生成对抗网络在人脸图像生成和编辑、风格迁移、图像超分辨率、图像修复,序列数据生成、视频生成等各个应用领域的算法以及对应算法的优缺点;介绍了生成对抗网络的常用评价指标并且分析了这些指标的适用场景和不足之处;最后从多个方面对生成对抗网络所面临的挑战进行了讨论,并指出了对其可能的改进方向。     

基于语义特征传播图神经网络的小样本图像分类算法

使用少量样本进行学习和概括的能力是人工智能和人类之间主要的区别。在小样本学习领域,大多数图神经网络专注于将标记的样本信息传递给未标记的查询样本,而忽略了语义特征在分类过程中的重要作用。为此构建了语义特征传播图神经网络,首先将语义特征嵌入到图神经网络中,解决了细粒度图像特征相似性带来的分类准确率低的问题,然后将注意力机制与骨干网络合并达到强化前景并提高特征提取质量的目的,利用马氏距离计算类的相似度得到更好的分类性能,最后使用Funnel ReLU函数作为激活函数进一步提高分类准确率。在基准数据集上实验表明,所提算法相比于基线算法在5类1/2/5样本任务上的准确率分别提高了903%、456%和415%。     

Dolphin Vocal Sound Generation via Deep WaveGAN

The marine biological sonar system evolved in the struggle of nature is far superior to the current artificial sonar. Therefore, the development of bionic underwater concealed detection is of great strategic significance to the military and economy. In this paper, a generative adversarial network (GAN) is trained based on the dolphin vocal sound dataset we constructed, which can achieve unsupervised generation of dolphin vocal sounds with global consistency. Through the analysis of the generated audio samples and the real audio samples in the time domain and the frequency domain, it can be proven that the generated audio samples are close to the real audio samples, which meets the requirements of bionic underwater concealed detection.     

Noise robust chi-square generative adversarial network

Aiming at the obvious difference of image quality generated by generative adversarial network under different noises,a chi-square generative adversarial network (CSGAN) was proposed.Combing the advantages of quantification sensitivity and sparse invariance,the chi-square divergence was introduced to calculate the distance between the generated samples and the original samples,which could reduce the influence of different noises on the generated samples and the quality requirement of original samples.Meanwhile,the network architecture was built and the global optimization objective function was constructed to enhance the adversarial performance.Experimental results show that the quality of the images generated by the proposed algorithm has little difference,and the network is more robust to different noises than the state-of-the-art networks.The application of chi-square divergence not only improves the quality of generated images,but also increases the robustness of the network under different noises.     

基于自监督对比学习的深度神经网络对抗鲁棒性提升

基于深度神经网络的多源图像内容自动分析与目标识别方法近年来不断取得新的突破,并逐步在智能安防、医疗影像辅助诊断和自动驾驶等多个领域得到广泛部署。然而深度神经网络的对抗脆弱性给其在安全敏感领域的部署带来巨大安全隐患。对抗鲁棒性的有效提升方法是采用最大化网络损失的对抗样本重训练深度网络,但是现有的对抗训练过程生成对抗样本时需要类别标记信息,并且会大大降低无攻击数据集上的泛化性能。本文提出一种基于自监督对比学习的深度神经网络对抗鲁棒性提升方法,充分利用大量存在的无标记数据改善模型在对抗场景中的预测稳定性和泛化性。采用孪生网络架构,最大化训练样本与其无监督对抗样本间的多隐层表征相似性,增强模型的内在鲁棒性。本文所提方法可以用于预训练模型的鲁棒性提升,也可以与对抗训练相结合最大化模型的“预训练+微调”鲁棒性,在遥感图像场景分类数据集上的实验结果证明了所提方法的有效性和灵活性。      

Adaptive adversarial prototyping network for few-shot prototypical translation

Translating multiple real-world source images to a single prototypical image is a challenging problem. Notably, these source images belong to unseen categories that did not exist during model training. We address this problem by proposing an adaptive adversarial prototype network (AAPN) and enhancing existing one-shot classification techniques. To overcome the limitations that traditional works cannot extract samples from novel categories, our method tends to solve the image translation task of unseen categories through a meta-learner. We train the model in an adversarial learning manner and introduce a style encoder to guide the model with an initial target style. The encoded style latent code enhances the performance of the network with conditional target style images. The AAPN outperforms the state-of-the-art methods in one-shot classification of brand logo dataset and achieves the competitive accuracy in the traffic sign dataset. Additionally, our model improves the visual quality of the reconstructed prototypes in unseen categories. Based on the qualitative and quantitative analysis, the effectiveness of our model for few-shot classification and generation is demonstrated.     

基于谱约束的生成对抗网络图像数据生成研究

数据匮乏是深度学习面临的一大难题。利用生成对抗网络(GAN)能够基于语义生成新的图像数据这一特性,提出一种基于谱约束的生成对抗网络图像数据生成方法,该方法针对卷积生成对抗网络模型易崩溃不收敛的问题,从每层神经网络的参数矩阵W的谱范数角度出发,引入谱范数归一化网络参数矩阵,将网络梯度限制在固定范围内,减缓判别网络收敛速度,从而提高GAN的训练稳定性。实验表明,通过该方法生成的数据相比原始GAN以及DCGAN、WGAN等生成的图像样本数据在图像识别网络中具有更高的准确率,能够对少量样本数据进行有效扩充。     

Feature discovering for image classification via wavelet-like pattern decomposition

In this paper, we propose a feature discovering method incorporated with a wavelet-like pattern decomposition strategy to address the image classification problem. In each level, we design a discriminative feature discovering dictionary learning (DFDDL) model to exploit the representative visual samples from each class and further decompose the commonality and individuality visual patterns simultaneously. The representative samples reflect the discriminative visual cues per class, which are beneficial for the classification task. Furthermore, the commonality visual elements capture the communal visual patterns across all classes. Meanwhile, the class-specific discriminative information can be collected by the learned individuality visual elements. To further discover the more discriminative feature information from each class, we then integrate the DFDDL into a wavelet-like hierarchical architecture. Due to the designed hierarchical strategy, the discriminative power of feature representation can be promoted. In the experiment, the effectiveness of proposed method is verified on the challenging public datasets.     

基于生成对抗网络的差分隐私数据发布方法

机器学习的飞速发展使其成为数据挖掘领域最有效的工具之一,但算法的训练过程往往需要大量的用户数据,给用户带来了极大的隐私泄漏风险.由于数据统计特征的复杂性及语义丰富性,传统隐私数据发布方法往往需要对原始数据进行过度清洗,导致数据可用性低而难以再适用于数据挖掘任务.为此,提出了一种基于生成对抗网络（Generative Adversarial Network,GAN）的差分隐私数据发布方法,通过在GAN模型训练的梯度上添加精心设计的噪声来实现差分隐私,确保GAN可无限量生成符合源数据统计特性且不泄露隐私的合成数据.针对现有同类方法合成数据质量低、模型收敛缓慢等问题,设计多种优化策略来灵活调整隐私预算分配并减小总体噪声规模,同时从理论上证明了合成数据严格满足差分隐私特性.在公开数据集上与现有方法进行实验对比,结果表明本方法能够更高效地生成质量更高的隐私保护数据,适用于多种数据分析任务.     

利用生成对抗网络的时频图像去噪和增强处理

针对雷达信号时频图像的去噪和增强问题,提出了利用生成对抗网络二次生成时频图像的方法。首先利用时频分析产生雷达信号的时频图像作为原始数据集1;接着利用生成对抗网络对数据集1进行学习之后生成新的数据集2,数据集2相对于数据集1拥有着去噪和增强的效果;最后提取时频图像奇异值特征检验生成的数据集2的有效性。对6种常见的雷达信号的时频图像进行了仿真实验,结果证明了该方法在时频图像去噪和增加样本多样性方面是有效的。