Joint image-to-image translation with denoising using enhanced generative adversarial networks

Impressive progress has been made recently in image-to-image translation using generative adversarial networks (GANs). However, existing methods often fail in translating source images with noise to target domain. To address this problem, we joint image-to-image translation with image denoising and propose an enhanced generative adversarial network (EGAN). In particular, built upon pix2pix, we introduce residual blocks in the generator network to capture deeper multi-level information between source and target image distribution. Moreover, a perceptual loss is proposed to enhance the performance of image-to-image translation. As demonstrated through extensive experiments, our proposed EGAN can alleviate effects of noise in source images, and outperform other state-of-the-art methods significantly. Furthermore, we experimentally indicate that the proposed EGAN is also effective when applied to image denoising.     

基于改进循环生成式对抗网络的图像风格迁移

图像间的风格迁移是一类将图片在不同领域进行转换的方法。随着生成式对抗网络在深度学习中的快速发展,其在图像风格迁移领域中的应用被日益关注。但经典算法存在配对训练数据较难获取,生成图片效果差的缺点。该文提出一种改进循环生成式对抗网络(CycleGAN++),取消了环形网络,并在图像生成阶段将目标域与源域的先验信息与相应图片进行纵深级联;优化了损失函数,采用分类损失代替循环一致损失,实现了不依赖训练数据映射的图像风格迁移。采用CelebA和Cityscapes数据集进行实验评测,结果表明在亚马逊劳务平台感知研究(AMT perceptual studies)与全卷积网络得分(FCN score)两个经典测试指标中,该文算法比CycleGAN, IcGAN, CoGAN, DIAT等经典算法取得了更高的精度。     

Learning discriminative and meaningful samples for generalized zero shot classification

Generalized zero shot classification aims to recognize both seen and unseen samples in test sets, which has gained great attention. Recently, many works consider using generative adversarial network to generate unseen samples for solving generalized zero shot classification problem. In this paper, we study how to generate discriminative and meaningful samples. We propose a method to learn discriminative and meaningful samples for generalized zero shot classification tasks (LDMS) by generative adversarial network with the regularization of class consistency and semantic consistency. In order to make the generated samples discriminative, class consistency is used, such that the generated samples of the same classes are near and of different classes are far away. In order to make the generated samples meaningful, semantic consistency is used, such that the semantic representations of the generated samples are close to their class prototypes. It encodes the discriminative information and semantic information to the generator. In order to alleviate the bias problem, we select some confident unseen samples. We use the seen samples, the generated unseen samples and the selected confident unseen samples to train the final classifier. Extensive experiments on all datasets demonstrate that the proposed method can outperform state-of-the-art models on generalized zero shot classification tasks.     

基于动态感受野的自适应多尺度信息融合的图像转换

为提高图像转换模型生成图像的质量,该文针对转换模型中的生成器进行改进,同时探究多样化的图像转换,拓展转换模型的生成能力。在生成器的改进方面,利用选择性(卷积)核模块(SKBlock)的动态感受野机制获取和融合生成器中每个上采样特征的多尺度信息,借助特征的多尺度信息和动态感受野构造选择性(卷积)核的生成式对抗网络(SK-GAN)。与传统生成器相比,SK-GAN以动态感受野获取多尺度信息的生成结构提高了生成图像的质量。在多样化图像转换方面,基于SK-GAN在草图合成真实图像任务提出带引导图像的选择性(卷积)核的生成式对抗网络(GSK-GAN)。该模型利用引导图像指导源图像的转换,通过引导图像编码器提取引导图像特征,然后由参数生成器(PG)和特征转换层(FT)将引导图像特征的信息传递至生成器。此外,该文还提出双分支引导图像编码器以提高转换模型的编辑能力,以及利用引导图像的隐变量分布实现随机样式的图像生成。实验表明,改进后的生成器有助于提高生成图像质量,SK-GAN在多个数据集中获得合理的生成结果。GSK-GAN不仅保证了生成图像的质量,还能生成更多样式的图像。     

A Semantics-based Translation Method for Automated Verification of SystemC TLM Designs

SystemC has become a de-facto standard language for SoC and ASIP designs. The verification of implementation with SystemC is the key to guarantee the correctness of designs and prevent the errors from propagating to the lower levels. In this project, we attempt translate SystemC programs to formal models and use existing model checkers to implement the verification. The method we proposed is based on a semantic translation method which translates sequential execution statements described as software character to parallel execution ones which are more closely with the implementation of hardware. This kind of conversion is inevitable to verify hardware designs but is overlooked in related works. The main contribution of this work is a translation method which can preserve the semantic consistency while building SMV model for SystemC design. We present the translation rules and implement a prototype tool which supports a subset of SystemC to demonstrate the effectiveness of our method.     

A comprehensive framework of multiple semantics preservation in neural style transfer

Recently neural style transfer has achieved great development, but there is still a big gap compared with manual creation. Most of the existing methods ignore the comprehensive consideration of preserving various semantic information of original content images, resulting in distortion or loss of original content features of the generated works, which are dull and difficult to convey the original themes and emotions. In this paper, we analyze the ability of the existing methods to maintain single semantic information and propose a fast style transfer framework with multi-semantic preservation. The experiments indicate that our method can effectively retain the original semantic information including salience and depth features, so that the final artwork has better visual effect by highlighting its regional focus and depth information. Compared with existing methods, our method has better ability in semantic preservation and can generate more artworks with distinct regions, controllable semantics, diverse contents and rich emotions.     

基于跨模态对比的场景图图像生成

条件图像生成根据不同形式的输入生成符合条件的图像,其中场景图是一类具有代表性的条件输入形式。场景图将图像中的物体抽象为节点,将物体之间的关系抽象为边,是一种广泛应用在计算机视觉和跨模态领域的结构化图表示。由于场景图中包含多个物体和物体之间的关系,现有的场景图图像生成方法容易导致生成结果和条件语义不一致,例如物体缺失和关系错误等。本文提出基于跨模态对比的生成方法解决上述问题。首先,本文提出关系一致性对比使生成的物体关系和输入的边保持一致。我们设计了联合特征代表图像中的物体的关系,并拉近联合特征和与其相关的边特征的距离,使其相比于不相关的边特征距离更接近。本文引入物体一致性对比使的生成的物体区域和输入的节点保持对应。在这个部分我们使用注意力机制获得节点对应的物体特征,然后拉近相关的节点特征于物体特征的距离。最后,本文提出全局一致性对比使的生成的图像整体和输入的场景图保持一致, 该对比损失将相关联的图像和场景图特征拉近,同时将不相关的样本特征相互远离。我们COCO-stuff和VG数据集上进行了详细的实验,实验结果表明我们的方法相比当前最佳性能分别在两个数据集上提升8.33%和8.87%的FID。消融实验表明每个对比损失模块都能够提升图像的生成质量,可视化结果展示了方法对于解决上述问题的有效性。从实验结果可知,我们的方法不仅能够提升图像的生成质量,并能够有效缓解物体缺失和关系错误等语义不一致问题。      

Single image dehazing using improved cycleGAN

Haze is an aggregation of very fine, widely dispersed, solid and/or liquid particles suspended in the atmosphere. In this paper, we propose an end-to-end network for single image dehazing, which enhances the CycleGAN model by introducing a transformer architecture within the generator, which is specific for haze removal. The proposed model is trained in an unpaired fashion with clear and hazy images altogether and does not require pairs of hazy and corresponding ground-truth clear images. Furthermore, the proposed model does not depend on estimating the parameters of the atmospheric scattering model. Rather, it uses a K-estimation module as the generator’s transformer for complete end-to-end modeling. The feature transformer introduced in the proposed generator model transforms the encoded features into desired feature space and then feeds them into the CycleGAN decoder to create a clear image. In the proposed model we further modified the cycle consistency loss to include the SSIM loss along with pixel-wise mean loss to produce a new loss function specific for the reconstruction task, which enhances the performance of the proposed model. The model performs well even on the high-resolution images provided in the NTIRE 2019 challenge dataset for single image dehazing. Further, we perform experiments on NYU-Depth and reside beta datasets. Results of our experiments show the efficacy of the proposed approach compared to the state-of-the-art in removing the haze from the input image.     

Clustering results of image searches by annotations and visual features

Image on web has become one of the most important information for browsers; however, the large number of results retrieved from images search engine increases the difficulty in finding the intended images. Image search result clustering (ISRC) is a solution to this problem. Currently, the ISRC-based methods separately utilized textual and visual features to present clustering result. In this paper, we proposed a new ISRC method as called Incremental-Annotations-based image search with clustering (IAISC), which adopted annotation as textual features and category model as visual features. IAISC can provide clustering result based on the semantic meaning and visual trail; further, presented by the iteratively structure, a user can obtain the intended image easily. The experimental result shows our method has high precision that the average precision rate is 73.4%; particularly, the precision rate is 96.5% when the user drills down the intended images till the last round. Regarding efficiency, our system is one and a half times as efficient as the previous studies.     

Fuzzy VA-Files for multi-label image annotation based on visual content of regions

In this work, we propose an efficient image annotation approach based on visual content of regions. We assume that regions can be described using low-level features as well as high-level ones. Indeed, given a labeled dataset, we adopt a probabilistic semantic model to capture relationships between low-level features and semantic clusters of regions. Moreover, since most previous works on image annotation do not deal with the curse of dimensionality, we solve this problem by introducing a fuzzy version of the Vector Approximation Files (VA-Files). Indeed, the main contribution of this work resides in the association of the generative model with fuzzy VA-Files, which offer an accurate multi-dimensional indexing, to estimate relationships between low-level features and semantic concepts. In fact, the proposed approach reduces the computation complexity while optimizing the annotation quality. Preliminary experiments highlight that the suggested approach outperforms other state-of-the-art approaches.     

Single image deraining with dual U-Net generative adversarial network

Most of the existing deraining methods cannot preserve the details of the image while removing the rain streaks. To solve this problem, we propose a single image de-raining method with dual U-Net generative adversarial network (DU-GAN). By using two U-Net with stronger learning ability as our generator DU-GAN can not only accurately remove more rain streaks but also preserve image details. The network can make full use of image information and extract complete image features. The adversarial loss function using the proposed dual U-Net generator is utilized to generate de-rained images which are close to the ground truth. Furthermore, to obtain the better visual effects of the generated image, The L1 and structure similarity loss functions which are consistent with the human visual effect are applied to generate the final output. The synthetic rainy image datasets and real rainy image datasets are used to evaluate the effectiveness of the proposed network in the experiments. The quantitative and visual experimental results show that the proposed single image deraining method achieves state-of-the-art compared with the other single image deraining methods. The source code can be found at https://github.com/LuBei-design/DU-GAN.

     

Ensemble similarity measure for community-based question answer简

Community-based question answer（CQA） makes a figure network in development of social network. Similar question retrieval is one of the most important tasks in CQA. Most of the previous works on similar question retrieval were given with the underlying assumption that answers are similar if their questions are similar, but no work was done by modeling similarity measure with the constraint of the assumption. A new method of modeling similarity measure is proposed by constraining the measure with the assumption, and employing ensemble learning to get a comprehensive measure which integrates different context features for similarity measuring, including lexical, syntactic, semantic and latent semantic. Experiments indicate that the integrated model could get a relatively high performance consistence between question set and answer set. Models with better consistency tend to get a better precision according to answers.     

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.     

ACGAN: Attribute controllable person image synthesis GAN for pose transfer

At present, pose transfer and attribute control tasks are still the challenges for image synthesis network. At the same time, there are often artifacts in the images generated by the image synthesis network when the above two tasks are completed. The existence of artifacts causes the loss of the generated image details or introduces some wrong image information, which leads to the decline of the overall performance of the existing work. In this paper, a generative adversarial network (GAN) named ACGAN is proposed to accomplish the above two tasks and effectively eliminate artifacts in generated images. The proposed network was compared quantitatively and qualitatively with previous works on the DeepFashion dataset and better results are obtained. Moreover, the overall network has advantages over the previous works in speed and number of parameters.     

Deep multi query image retrieval

There exist few studies investigating the multi-query image retrieval problem. Existing methods are not based on hash codes. As a result, they are not efficient and fast. In this study, we develop an efficient and fast multi-query image retrieval method when the queries are related to more than one semantic. Image hash codes are generated by a deep hashing method. Consequently, the method requires lower storage space, and it is faster compared to the existing methods. The retrieval is based on the Pareto front method. Reranking performed on the retrieved images by using non-binary deep-convolutional features increase retrieval accuracy considerably. Unlike previous studies, the method supports an arbitrary number of queries. It outperforms similar multi-query image retrieval studies in terms of retrieval time and retrieval accuracy.     

AMSFF-Net: Attention-Based Multi-Stream Feature Fusion Network for Single Image Dehazing

In this paper, an end-to-end convolutional neural network is proposed to recover haze-free image named as Attention-Based Multi-Stream Feature Fusion Network (AMSFF-Net). The encoder-decoder network structure is used to construct the network. An encoder generates features at three resolution levels. The multi-stream features are extracted using residual dense blocks and fused by feature fusion blocks. AMSFF-Net has ability to pay more attention to informative features at different resolution levels using pixel attention mechanism. A sharp image can be recovered by the good kernel estimation. Further, AMSFF-Net has ability to capture semantic and sharp textural details from the extracted features and retain high-quality image from coarse-to-fine using mixed-convolution attention mechanism at decoder. The skip connections decrease the loss of image details from the larger receptive fields. Moreover, deep semantic loss function emphasizes more semantic information in deep features. Experimental findings prove that the proposed method outperforms in synthetic and real-world images.     

Semantic prior-driven fused contextual transformation network for image inpainting

Recent advances in image inpainting have achieved impressive performance for generating plausible visual details on small regular image defects or simple backgrounds. However, current solution suffers from the lack of semantic priors for the image and the inability to deduce the image content from distant background, leading to distorted structures and artifacts in the results when inpainting large random irregular complicated images. To address these problems, a semantic prior-driven fused contextual transformation network for image inpainting is proposed as a promise solution. First, the semantic prior generator is put forward to map the semantic features of ground truth images and the low-level features of broken images to semantic priors. Subsequently, an image split-transform-aggregated strategy, named fusion context transformation block, is presented to infer rich multi-scale remote texture features and thus to improve the restored image finesse. Thereafter, an aggregated semantic attention-aware module, consisting of spatially adaptive normalization and enhanced spatial attention is designed to aggregate semantic priors and multi-scale texture features into the decoder to restore reasonable structure. Finally, the mask guided discriminator is developed to effectively discriminate between real and false pixels in the output image to improve the capability of the discriminator and hence to reduce the probability of artifacts containing in the output image. Comprehensive experimental results on CelebA-HQ, Paris Street View, and Places2 datasets demonstrate the superiority of the proposed network over the state-of-the-arts, whose PSNR, SSIM and MAE are improved about 20 %, 12.6 %, and 42 % gains, respectively.     

面向大规模社区问答数据的问题检索方法

随着问答社区网站的兴起,越来越多的用户生成数据积累了起来。这些用户生成数据不仅具有海量的、多样性的等特点,还有着极高的质量和重用价值。为了高效地管理和利用这些数据,近年来研究人员基于这些数据进行了大量的研究和实践,而社区问答中的问题检索就是一个被广泛研究的课题。主要研究了面向大规模社区问答数据的问题检索方法。收集来自Yahoo!Answers等社区网站的超过1．3亿问题和10亿答案的大规模数据,与之前的基于百万量级的数据的问答社区相关研究工作相比有着明显的不同和极高的实用价值。在此数据的基础上,通过查询自动分类方法来提高每次查询效率和效果。在问题检索过程中,提出了应用查询问句和问题的结构信息和语义信息,结合排序学习算法来融合多种不同类别的特征的方法,通过应用训练数据生成排序模型来提高问题检索的相关性和词语不匹配等问题。实验表明,本文应用RankingSVM方法来训练的排序模型在不同数据集上,其准确率等评价指标上都相比以往的方法有着显著的提高。     

基于纹元森林和显著性先验的弱监督图像语义分割方法

弱监督语义分割任务常利用训练集中全体图像的超像素及其相似度建立图模型,使用图像级别标记的监督关系进行约束求解。全局建模缺少单幅图像结构信息,同时此类参数方法受到复杂度限制,无法使用大规模的弱监督训练数据。针对以上问题,该文提出一种基于纹元森林和显著性先验的弱监督图像语义分割方法。算法使用弱监督数据和图像显著性训练随机森林分类器用于语义纹元森林特征(Semantic Texton Forest, STF)的提取。测试时,先将图像进行过分割,然后提取超像素语义纹元特征,利用朴素贝叶斯法进行超像素标记的概率估计,最后在条件随机场(CRF)框架下结合图像显著性信息定义了新的能量函数表达式,将图像的标注(labeling)问题转换为能量最小化问题求解。在MSRC-21类数据库上进行了验证,完成了语义分割任务。结果表明,在并未对整个训练集建立图模型的情况下,仅利用单幅图像的显著性信息也可以得到较好的分割结果,同时非参模型有利于规模数据分析。