基于相对生成对抗网络的低清小脸幻构

针对当前代表性低清小脸幻构方法存在的视觉真实感弱、网络结构复杂等问题,提出了一种基于相对生成对抗网络的低清小脸幻构方法(tfh-RGAN).该文方法的网络架构包括幻构生成器和判别器两个部分,通过像素损失函数和相对生成对抗损失函数的联合最小化,实现生成器和判别器的交替迭代训练.其中,幻构生成器结合了残差块、稠密块以及深度可分离卷积算子,保证幻构效果和网络深度的同时降低生成器的参数量;判别器采用图像分类问题中的全卷积网络,通过先后去除批归一化层、添加全连接层,充分挖掘相对生成对抗网络在低清小脸幻构问题上的能力极限.实验结果表明,在不额外显式引入任何人脸结构先验的条件下,该文方法能够以更简练的网络架构输出清晰度更高、真实感更强的幻构人脸.从定量角度看,该文方法的峰值信噪比相较之前的若干代表性方法可提高0.25～1.51 dB.     

基于生成对抗网络的图像转换技术

针对不同谱段图像获取代价不同的问题,提出一种基于生成对抗网络的图像转换方法。转换过程以肉眼可分辨范围内图像轮廓不变为出发点。首先,通过成对的训练数据对生成器和判别器进行交替训练,不断对损失函数进行优化,直到模型达到纳什平衡。然后用测试数据对上述训练好的模型进行检测,查看转换效果,并从主观观察和客观上计算平均绝对误差和均方误差角度评价转换效果。通过上述过程最终实现不同谱段图像之间的转换。其中,生成器借鉴U-Net架构;判别器采用传统卷积神经网络架构;损失函数方面增加L1损失来保证图像转换前后高、低频特征的完整性。以红外图像与可见光图像之间的转换为例进行实验,结果表明,通过本文设计的生成对抗网络,可以较好地实现红外图像与可见光图像之间的转换。     

基于改进深度卷积生成对抗网络的刺绣图像修复

目前中华传统刺绣工艺传承保护问题中的修复任务主要以人工为主，修复过程需要大量的人力、物力。随着深度学习的高速发展，不同类型的刺绣文物损伤可以利用生成对抗网络进行修复。针对上述问题，提出一种基于改进深度卷积生成对抗网络（DCGAN）的刺绣图像修复方法。首先，在生成器部分引入空洞卷积层扩大感受野，并添加卷积注意力机制模块，在通道与空间2个维度增强重要特征的指导作用；在判别器部分增加全连接层数提升网络解决非线性问题的能力；在损失函数部分联合均方误差损失与对抗损失通过网络训练相互博弈的过程实现刺绣图像修复。实验结果表明：引入空洞卷积层与注意力机制提升了网络性能与修复效果，最终得到修复图像的结构相似性高达0.955，能够得到较为自然的刺绣图像修复效果，可以为专家提供纹理、色彩等信息作为参考辅助后续的修复。     

属性一致的物体轮廓划分模型

该文提出一种基于全卷积深度残差网络、结合生成式对抗网络思想的基于属性一致的物体轮廓划分模型。采用物体轮廓划分网络作为生成器进行物体轮廓划分;该网络运用结构相似性作为区域划分的重构损失,从视觉系统的角度监督指导模型学习;使用全局和局部上下文判别网络作为双路判别器,对区域划分结果进行真伪判别的同时,结合对抗式损失提出一种联合损失用于监督模型的训练,使区域划分内容真实、自然且具有属性一致性。通过实例验证了该方法的实时性、有效性。     

属性一致的物体轮廓划分模型

该文提出一种基于全卷积深度残差网络、结合生成式对抗网络思想的基于属性一致的物体轮廓划分模型.采用物体轮廓划分网络作为生成器进行物体轮廓划分;该网络运用结构相似性作为区域划分的重构损失,从视觉系统的角度监督指导模型学习;使用全局和局部上下文判别网络作为双路判别器,对区域划分结果进行真伪判别的同时,结合对抗式损失提出一种联合损失用于监督模型的训练,使区域划分内容真实、自然且具有属性一致性.通过实例验证了该方法的实时性、有效性.     

改进生成式对抗网络的图像数据集增强算法

针对现有深度学习中图像数据集缺乏的问题,提出了一种基于深度卷积生成式对抗网络(Deep Convolutional Generative Adversarial Network, DCGAN)的图像数据集增强算法。该算法对DCGAN网络进行改进,首先在不过多增加计算量的前提下改进现有的激活函数,增强生成特征的丰富性与多样性;然后通过引入相对判别器有效缓解模式坍塌现象,从而提升模型稳定性;最后在现有生成器结构中引入残差块,获得相对高分辨率的生成图像。实验结果表明,将所提方法应用在MNIST、SAR和医学血细胞数据集上,图像数据增强效果与未改进的DCGAN网络相比显著提升。     

基于改进生成对抗网络模型的红外与可见光图像融合

为增强融合图像的视觉效果,减少计算的复杂度,解决传统红外与可见光图像融合算法存在的背景细节丢失问题,提出了一种生成对抗网络框架下基于深度可分离卷积的红外与可见光图像融合方法。首先,在生成器中对源图像进行深度卷积与逐点卷积运算,得到源图像的特征映射信息;其次,通过前向传播的方式更新网络参数,得到初步的单通道融合图像;再次,在红外及可见光判别器中,使用深度可分离卷积分别对源图像与初步融合图像进行像素判别;最后,在损失函数的约束下,双判别器不断将更多的细节信息添加到融合图像中。实验结果表明,相比于传统的融合算法,该方法在信息熵、平均梯度、空间频率、标准差、结构相似性损失和峰值信噪比等评价指标上分别平均提高了1.63%、1.02%、3.54%、5.49%、1.05%、0.23%,在一定程度上提升了融合图像的质量,丰富了背景的细节信息。     

红外与可见光图像注意力生成对抗融合方法研究

目前,基于深度学习的融合方法依赖卷积核提取局部特征,而单尺度网络、卷积核大小以及网络深度的限制无法满足图像的多尺度与全局特性.为此,本文提出了红外与可见光图像注意力生成对抗融合方法.该方法采用编码器和解码器构成的生成器以及两个判别器.在编码器中设计了多尺度模块与通道自注意力机制,可以有效提取多尺度特征,并建立特征通道长...     

基于多通道GAN的图像去噪算法

针对图像采集和传输过程中所产生噪声导致后续图像处理能力下降的问题,提出基于生成对抗网络(GAN)的多通道图像去噪算法。所提算法将含噪彩色图像分离为RGB三通道,各通道基于具有相同架构的端到端可训练的GAN实现去噪。GAN生成网络基于U-net衍生网络以及残差块构建,从而可参考低级特征信息以有效提取深度特征进而避免丢失细节信息;判别网络则基于全卷积网络构造,因而可获得像素级分类从而提升判别精确性。此外,为改善去噪能力且尽可能保留图像细节信息,所构建去噪网络基于对抗损失、视觉感知损失和均方误差损失这3类损失度量构建复合损失函数。最后,利用算术平均方法融合三通道输出信息以获得最终去噪图像。实验结果表明,与主流算法相比,所提算法可有效去除图像噪声,且可较好地恢复原始图像细节。     

基于DCGAN的数据增强方法

针对深度学习中训练样本数量少，难以满足深度学习需求的问题，提出一种基于深度卷积生成对抗网络（Deep Convolutional Generative Adversarial Networks,DCGAN）的数据增强方法，将深度学习运用于数据增强，利用卷积神经网络（Convolutional Neural Network,CNN）构建生成器与鉴别器网络，引入高斯误差线性单元（Gaussian Error Linear Unit,GELU）作为激活函数。将名人头像数据集（Large-scale CelebFaces Attributes,Celeb A）作为训练样本进行训练。试验结果表明，该网络的鉴别器损失值快速收敛于0附近，生成器损失值快速收敛于3附近，收敛速度较快。网络能够根据学习到的样本数据分布，生成逼真的样本数据，该方法能够有效解决样本数量稀少的问题，实现数据增强。     

Hallucinating face by eigentransformation

In video surveillance, the faces of interest are often of small size. Image resolution is an important factor affecting face recognition by human and computer. In this paper, we propose a new face hallucination method using eigentransformation. Different from most of the proposed methods based on probabilistic models, this method views hallucination as a transformation between different image styles. We use Principal Component Analysis (PCA) to fit the input face image as a linear combination of the low-resolution face images in the training set. The high-resolution image is rendered by replacing the low-resolution training images with high-resolution ones, while retaining the same combination coefficients. Experiments show that the hallucinated face images are not only very helpful for recognition by humans, but also make the automatic recognition procedure easier, since they emphasize the face difference by adding more high-frequency details.     

Generative adversarial networks for hyperspectral image spatial super-resolution

It is becoming increasingly easier to obtain more abundant supplies for hyperspectral images ( HSIs). Despite this, achieving high resolution is still critical. In this paper, a method named hyperspectral images super-resolution generative adversarial network ( HSI-RGAN ) is proposed to enhance the spatial resolution of HSI without decreasing its spectral resolution. Different from existing methods with the same purpose, which are based on convolutional neural networks ( CNNs) and driven by a pixel-level loss function, the new generative adversarial network (GAN) has a redesigned framework and a targeted loss function. Specifically, the discriminator uses the structure of the relativistic discriminator, which provides feedback on how much the generated HSI looks like the ground truth. The generator achieves more authentic details and textures by removing the place of the pooling layer and the batch normalization layer and presenting smaller filter size and two-step upsampling layers. Furthermore, the loss function is improved to specially take spectral distinctions into account to avoid artifacts and minimize potential spectral distortion, which may be introduced by neural networks. Furthermore, pre-training with the visual geometry group (VGG) network helps the entire model to initialize more easily. Benefiting from these changes, the proposed method obtains significant advantages compared to the original GAN. Experimental results also reveal that the proposed method performs better than several state-of-the-art methods.     

Generative adversarial network-based atmospheric scattering model for image dehazing

This paper presents a trainable Generative Adversarial Network (GAN)-based end-to-end system for image dehazing, which is named the DehazeGAN. DehazeGAN can be used for edge computing-based applications, such as roadside monitoring. It adopts two networks: one is generator (G), and the other is discriminator (D). The G adopts the U-Net architecture, whose layers are particularly designed to incorporate the atmospheric scattering model of image dehazing. By using a reformulated atmospheric scattering model, the weights of the generator network are initialized by the coarse transmission map, and the biases are adaptively adjusted by using the previous round’s trained weights. Since the details may be blurry after the fog is removed, the contrast loss is added to enhance the visibility actively. Aside from the typical GAN adversarial loss, the pixel-wise Mean Square Error (MSE) loss, the contrast loss and the dark channel loss are introduced into the generator loss function. Extensive experiments on benchmark images, the results of which are compared with those of several state-of-the-art methods, demonstrate that the proposed DehazeGAN performs better and is more effective.     

High-quality face image generation based on generative adversarial networks

Conventional face image generation using generative adversarial networks (GAN) is limited by the quality of generated images since generator and discriminator use the same backpropagation network. In this paper, we discuss algorithms that can improve the quality of generated images, that is, high-quality face image generation. In order to achieve stability of network, we replace MLP with convolutional neural network (CNN) and remove pooling layers. We conduct comprehensive experiments on LFW, CelebA datasets and experimental results show the effectiveness of our proposed method.     

Cross-view information interaction and feedback network for face hallucination

Hallucinating a photo-realistic frontal face image from a low-resolution (LR) non-frontal face image is beneficial for a series of face-related applications. However, previous efforts either focus on super-resolving high-resolution (HR) face images from nearly frontal LR counterparts or frontalizing non-frontal HR faces. It is necessary to address all these challenges jointly for real-world face images in unconstrained environment. In this paper, we develop a novel Cross-view Information Interaction and Feedback Network (CVIFNet), which simultaneously handles the non-frontal LR face image super-resolution (SR) and frontalization in a unified framework and interacts them with each other to further improve their performance. Specifically, the CVIFNet is composed of two feedback sub-networks for frontal and profile face images. Considering the reliable correspondence between frontal and non-frontal face images can be crucial and contribute to face hallucination in a different manner, we design a cross-view information interaction module (CVIM) to aggregate HR representations of different views produced by the SR and frontalization processes to generate finer face hallucination results. Besides, since 3D rendered facial priors contain rich hierarchical features, such as low-level (e.g., sharp edge and illumination) and perception level (e.g., identity) information, we design an identity-preserving consistency loss based on 3D rendered facial priors, which can ensure that the high-frequency details of frontal face hallucination result are consistent with the profile. Extensive experiments demonstrate the effectiveness and advancement of CVIFNet.     

角度空间三元组损失微调的人脸识别

针对角度Softmax损失强约束存在的问题,提出一种用角度空间三元组损失对角度Softmax损失预训练模型进行微调的算法。算法首先对原来的卷积神经网络结构进行改进,将1×1卷积核与池化层加在不同残差块间,以进行选择更有效的特征。然后用角度空间下的三元组损失对预训练模型进行微调,以降低困难样本的强约束条件。最后在测试时,分别提取原始人脸图像特征和水平翻转的人脸图像特征,对两个特征相加作为最终的人脸特征表达,以丰富人脸特征信息,提高识别率。实验结果表明,在LFW和YTF人脸数据集分别取得了99.25%和94.52%的识别率,在大规模人脸身份识别中,本文提出的方法在仅用单模型和比较小的训练集就能有效地提高人脸识别率。     

Generalized face super-resolution.

Existing learning-based face super-resolution (hallucination) techniques generate high-resolution images of a single facial modality (i.e., at a fixed expression, pose and illumination) given one or set of low-resolution face images as probe. Here, we present a generalized approach based on a hierarchical tensor (multilinear) space representation for hallucinating high-resolution face images across multiple modalities, achieving generalization to variations in expression and pose. In particular, we formulate a unified tensor which can be reduced to two parts: a global image-based tensor for modeling the mappings among different facial modalities, and a local patch-based multiresolution tensor for incorporating high-resolution image details. For realistic hallucination of unregistered low-resolution faces contained in raw images, we develop an automatic face alignment algorithm capable of pixel-wise alignment by iteratively warping the probing face to its projection in the space of training face images. Our experiments show not only performance superiority over existing benchmark face super-resolution techniques on single modal face hallucination, but also novelty of our approach in coping with multimodal hallucination and its robustness in automatic alignment under practical imaging conditions.     

融合注意力机制的高分辨人脸识别图像重建

针对由于人脸姿势、光照不均、拍摄环境、拍摄设备等内外部因素造成图像分辨率低的问题,提出融合注意力机制的高分辨人脸识别图像重建模型.首先以低分辨率人脸图像对作为两个生成器输入,通过残差块和注意力模块堆叠网络提取人脸特征信息,进而生成高分辨率人脸图像.训练中使用一个鉴别器来监督两个生成器的训练过程.利用Adam算法对鉴别器...     

Stable self-attention adversarial learning for semi-supervised semantic image segmentation

The application of adversarial learning for semi-supervised semantic image segmentation based on convolutional neural networks can effectively reduce the number of manually generated labels required in the training process. However, the convolution operator of the generator in the generative adversarial network (GAN) has a local receptive field, so that the long-range dependencies between different image regions can only be modeled after passing through multiple convolutional layers. The present work addresses this issue by introducing a self-attention mechanism in the generator of the GAN to effectively account for relationships between widely separated spatial regions of the input image with supervision based on pixel-level ground truth data. In addition, the adjustment of the discriminator has been demonstrated to affect the stability of GAN training performance. This is addressed by applying spectral normalization to the GAN discriminator during the training process. Our method has better performance than existing full/semi-supervised semantic image segmentation techniques.     

基于单一神经网络的多尺度人脸检测

人脸检测是指检测并定位输入图像中所有的人脸,并返回精确的人脸位置和大小,是目标检测的重要方向。为了解决人脸尺度多样性给人脸检测造成的困难,该文提出一种新的基于单一神经网络的特征图融合多尺度人脸检测算法。该算法在不同大小的卷积层上预测人脸,实现实时多尺度人脸检测,并通过将浅层的特征图融合引入上下文信息提高小尺寸人脸检测精度。在数据集FDDB和WIDERFACE测试结果表明,所提方法达到了先进人脸检测的水平,并且该方法去掉了框推荐过程,因此检测速度更快。在WIDERFACE难、适中、简单3个子数据集上测试结果分别为87.9%, 93.2%, 93.4% MAP,检测速度为35 fps。所提算法与目前效果较好的极小人脸检测方法相比,在保证精度的同时提高了人脸检测速度。