优化非线性激活函数-全局卷积神经网络的物体识别算法

传统的物体识别算法识别精度、自适应能力弱等问题已然不能满足实际的仓储物流领域对物体识别精度的要求.近年来,相关学者提出了基于深度学习的物体识别算法,它得到一定的推广和应用.但是,深度学习在物体识别的应用过程中存在以下问题:一是深度学习模型中激活函数的非线性建模能力弱;二是深度学习模型大量重复的池化操作丢失信息.鉴于此,本文提出了一种参数形式统一且可学习的指数非线性单元(Multiple Parameters Exponential Linear Units,MPELU).它通过在ELU(Exponential Linear Units)中引入两个学习的参数,提升模型的非线性建模能力.同时,本文提出了一种新的全局卷积神经网络结构,减少大量池化操作丢失特征信息的问题.基于上述思想,本文提出了优化非线性激活函数-全局卷积神经网络的物体识别算法.利用本文所提算法对CIFAR100数据集和ImageNet数据集分别进行实验.结果表明,本文所提物体识别方法不仅识别准确率较传统机器学习、其他深度学习模型有较大幅度提升,而且具有良好的稳定性和鲁棒性.     

Multiphase B-spline level set and incremental shape priors with applications to segmentation and tracking of left ventricle in cardiac MR images

This paper presents a new multiphase active contour model for object segmentation and tracking. The paper introduces an energy functional which incorporates image feature information to drive contours toward desired boundaries, and shape priors to constrain the evolution of the contours with respect to reference shapes. The shape priors, in the model, are constructed by performing the incremental principal component analysis (iPCA) on a set of training shapes and newly available shapes which are the resulted shapes derived from preceding segmented images. By performing iPCA, the shape priors are updated without repeatedly performing PCA on the entire training set including the existing shapes and the newly available shapes. In addition, by incrementally updating the resulted shape information of consecutive frames, the approach allows to encode shape priors even when the database of training shapes is not available. Moreover, in shape alignment steps, we exploit the shape normalization procedure, which takes into account the affine transformation, to directly calculate pose transformations instead of solving a set of coupled partial differential equations as in gradient descent-based approaches. Besides, we represent the level set functions as linear combinations of continuous basic functions expressed on B-spline basics for a fast convergence to the segmentation solution. The model is applied to simultaneously segment/track both the endocardium and epicardium of left ventricle from cardiac magnetic resonance (MR) images. Experimental results show the desired performances of the proposed model.     

Extraction of left ventricle borders with local and global priors from echocardiograms

This paper presents a novel technique for the extraction of the left ventricle borders from echocardiograms with prior information. Although the literature includes many successful prior based methods, priors that include both image and non-image related features are rare for the contour extraction. We classify these features as local and global priors where the local priors refer to the locally definable features of the target borders and global priors refer to the geometric shape properties. The local priors, which include image, motion, and local shape information, are learned with AdaBoost. The scores produced by AdaBoost for the target images are combined with the global shape prior under a level set framework. The main contributions of this paper are to learn different types of local features efficiently with machine learning and to combine these features with the geometric shape information for the contour extraction task. The system is validated on the real echocardiograms and synthetic images. The results indicate that using local and global priors together produces better extraction results and the contours extracted by the proposed system are in accord with the expert delineated borders.     

Salient object detection via local saliency estimation and global homogeneity refinement

This paper presents a new hybrid approach for detecting salient objects in an image. It consists of two processes: local saliency estimation and global-homogeneity refinement. We model the salient object detection problem as a region growing and competition process by propagating the influence of foreground and background seed-patches. First, the initial local saliency of each image patch is measured by fusing local contrasts with spatial priors, thereby the seed-patches of foreground and background are constructed. Later, the global-homogeneous information is utilized to refine the saliency results by evaluating the ratio of the foreground and background likelihoods propagated from the seed-patches. Despite the idea is simple, our method can effectively achieve consistent performance for detecting object saliency. The experimental results demonstrate that our proposed method can accomplish remarkable precision and recall rates with good computational efficiency.     

一种基于图卷积网络的文本多标签学习方法

多标签学习广泛应用于文本分类、标签推荐、主题标注等.最近,基于深度学习技术的多标签学习受到广泛关注,针对如何在多标签学习中有效挖掘并利用高阶标签关系的问题,提出一种基于图卷积网络探究标签高阶关系的模型TMLLGCN.该模型采用GCN的映射函数从数据驱动的标签表示中生成对象分类器挖掘标签高阶关系.首先,采用深度学习方法提取文本特征,然后以数据驱动方式获得基础标签关联表示矩阵,为更好地建模高阶关系及提高模型效果,在基础标签关联表示矩阵上考虑未标记标签集对已知标签集的影响进行标签补全,并以此相关性矩阵指导GCN中标签节点之间的信息传播,最后将提取的文本特征应用到学习高阶标签关系的图卷积网络分类器进行端到端训练,综合标签关联和特征信息作为最终的预测结果.在实际多标签数据集上的实验结果表明,提出的模型能够有效建模标签高阶关系且提升了多标签学习的效果.     

Object tracking using deformable templates

We propose a method for object tracking using prototype-based deformable template models. To track an object in an image sequence, we use a criterion which combines two terms: the frame-to-frame deviations of the object shape and the fidelity of the modeled shape to the input image. The deformable template model utilizes the prior shape information which is extracted from the previous frames along with a systematic shape deformation scheme to model the object shape in a new frame. The following image information is used in the tracking process: 1) edge and gradient information: the object boundary consists of pixels with large image gradient, 2) region consistency: the same object region possesses consistent color and texture throughout the sequence, and 3) interframe motion: the boundary of a moving object is characterized by large interframe motion. The tracking proceeds by optimizing an objective function which combines both the shape deformation and the fidelity of the modeled shape to the current image (in terms of gradient, texture, and interframe motion). The inherent structure in the deformable template, together with region, motion, and image gradient cues, makes the proposed algorithm relatively insensitive to the adverse effects of weak image features and moderate amounts of occlusion     

基于深度强化学习的遥感图像可解释目标检测方法

随着遥感技术的飞速发展,遥感图像目标检测在资源勘探、城市规划、自然灾害评估等方面得到广泛应用.遥感影像背景复杂、目标尺度较小,难以检测.针对此问题,文中提出基于深度强化学习的遥感图像可解释目标检测方法.首先,将深度强化学习应用于超快速区域神经网络中的候选区域生成网络,修改激励函数,提高对遥感图像的检测精度.然后,将原有参数量较大的主干网络轻量化,提高方法的检测速度和可移植性.最后,利用网络解剖方法对隐层表征的可解释性进行量化,赋予方法人类理解的可解释性概念.实验表明,文中方法在3个公开的遥感数据集上的性能有所提升.通过改进的网络解剖方法进一步验证方法的有效性.     

基于语义规则与RNN模型的在线评论情感分类研究

为提高互联网中在线评论文本的情感倾向分类准确率,方便消费者和商家准确高效地获取信息,该文提出一种将语义规则方法与深度学习方法相结合的在线评论文本情感分类模型,对基于情感词典的语义规则信息进行扩展,嵌入到常用特征模板中组合成更有效的混合特征模板;采用Fisher判别准则方法对混合特征模板进行降维以消除特征间的信息冗余;深度学习模型采用基于LSTM改进的RNN模型,将网络爬取的数据输入到模型进行训练和测试。结果表明,语义规则抽取出的特征包含更多、更准确的情感信息,使得混合特征模板可以更加全面地考虑文本的情感特征粒度;Fisher准则可有效识别出高判别性的低维文本特征,进一步提高改进RNN模型对评论文本的分类性能。     

GeoAI in terrain analysis: Enabling multi-source deep learning and data fusion for natural feature detection

In this paper we report on a new GeoAI research method which enables deep machine learning from multi-source geospatial data for natural feature detection. In particular, a multi-source, deep learning-based object detection pipeline was developed. This pipeline introduces three new features: First, strategies of both data-level fusion (i.e., channel expansion on convolutional neural networks) and feature-level fusion were integrated into the object detection model to allow simultaneous machine learning from multi-source data, including remote sensing imagery and Digital Elevation Model (DEM) data. Second, a new data fusion strategy was developed to blend DEM data and its derivatives to create a new, fused data source with enriched information content and image features. The model has also enabled deep learning by combining both the proposed data fusion and feature-level fusion strategies to yield a much-improved detection result. Third, two different sets of data augmentation techniques were applied to the multi-source training data to further improve the model performance. A series of experiments were conducted to verify the effectiveness of the proposed strategies in multi-source deep learning.     

全局对比和背景先验驱动的显著目标检测

针对传统背景先验方法中背景提取不精确并且背景抑制能力弱的问题,提出了全局对比和背景先验驱动的显著目标检测方法。首先将图像分割为一系列感知均匀的超像素,再由全局颜色对比得到基于全局的显著图并计算得到前景种子点;然后将每个边界超像素与前景种子点做对比,筛选差异性较大的边界超像素作为背景种子点并计算得到基于背景的显著图;最后在融合基于全局和背景显著图的基础上,提出一种多兴趣点高斯模型的方法进一步抑制背景并整体高亮显著区域。在公开的MSRA-1000数据测试集上与6种主流方法进行对比实验,结果表明,所提出的显著性目标检测方法对复杂边界信息具有更强的鲁棒性,并能有效抑制背景噪声。     

基于DDPG的三维重建模糊概率点推理

单视图物体三维重建是一个长期存在的具有挑战性的问题. 为了解决具有复杂拓扑结构的物体以及一些高保真度的表面细节信息仍然难以准确进行恢复的问题, 本文提出了一种基于深度强化学习算法深度确定性策略梯度 (Deep deterministic policy gradient, DDPG)的方法对三维重建中模糊概率点进行再推理, 实现了具有高保真和丰富细节的单视图三维重建. 本文的方法是端到端的, 包括以下四个部分: 拟合物体三维形状的动态分支代偿网络的学习过程, 聚合模糊概率点周围点的邻域路由机制, 注意力机制引导的信息聚合和基于深度强化学习算法的模糊概率调整. 本文在公开的大规模三维形状数据集上进行了大量的实验证明了本文方法的正确性和有效性. 本文提出的方法结合了强化学习和深度学习, 聚合了模糊概率点周围的局部信息和图像全局信息, 从而有效地提升了模型对复杂拓扑结构和高保真度的细节信息的重建能力.     

Iterated random walks with shape prior

We propose a new framework for image segmentation using random walks where a distance shape prior is combined with a region term. The shape prior is weighted by a confidence map to reduce the influence of the prior in high gradient areas and the region term is computed with k-means to estimate the parametric probability density function. Then, random walks is performed iteratively aligning the prior with the current segmentation in every iteration. We tested the proposed approach with natural and medical images and compared it with the latest techniques with random walks and shape priors. The experiments suggest that this method gives promising results for medical and natural images.     

金字塔频率特征融合目标检测网络

针对深度学习网络在特征提取过程中运用上采样操作而致使细节纹理等高频特征缺失的问题,提出一种金字塔频率特征融合目标检测网络.网络由3个深度学习金字塔网络构成,输入图像经初级金字塔提取深度特征后,分别通过高频、低频增强金字塔形成不同的频率特征,利用特征融合来凸显深度学习网络在信息逐层传递过程中对细节信息的保护能力,提高目标...     

结合语义边界信息的道路环境语义分割方法

语义分割是实现道路语义环境解释的重要方法,深度学习语义分割由于卷积、池化及反卷积的作用使分割边界模糊、不连续以及小目标漏分错分,影响了分割效果,降低了分割精度。针对上述问题,提出了一种结合语义边界信息的新的语义分割方法,首先在语义分割深度模型中构建了一个语义边界检测子网,利用网络中的特征共享层将语义边界检测子网络学习到的语义边界信息传递给语义分割网络;然后结合语义边界检测任务和语义分割任务定义了新的模型代价函数,同时完成语义边界检测和语义分割两个任务,提升语义分割网络对物体边界的描述能力,提高语义分割质量。最后在Cityscapes数据集上进行一系列实验证明,结合语义边界信息的语义分割方法在准确率上比已有的语义分割网络SegNet提升了2.9%,比ENet提升了1.3%。所提方法可以改善语义分割中出现的分割不连续、物体边界不清晰、小目标错分漏分、分割精度不高等问题。     

注意力引导网络的显著性目标检测

目的 全卷积模型的显著性目标检测大多通过不同层次特征的聚合实现检测,如何更好地提取和聚合特征是一个研究难点。常用的多层次特征融合策略有加法和级联法,但是这些方法忽略了不同卷积层的感受野大小以及产生的特征图对最后显著图的贡献差异等问题。为此,本文结合通道注意力机制和空间注意力机制有选择地逐步聚合深层和浅层的特征信息,更好地处理不同层次特征的传递和聚合,提出了新的显著性检测模型AGNet(attention-guided network),综合利用几种注意力机制对不同特征信息加权解决上述问题。方法 该网络主要由特征提取模块(feature extraction module, FEM)、通道—空间注意力融合模块(channel-spatial attention aggregation module, C-SAAM)和注意力残差细化模块(attention residual refinement module,ARRM)组成,并且通过最小化像素位置感知(pixel position aware, PPA)损失训练网络。其中,C-SAAM旨在有选择地聚合浅层的边缘信息以及深层抽象的语义特征,利用通道注意力和空间注意力避免融合冗余的背景信息对显著性映射造成影响;ARRM进一步细化融合后的输出,并增强下一个阶段的输入。结果 在5个公开数据集上的实验表明,AGNet在多个评价指标上达到最优性能。尤其在DUT-OMRON(Dalian University of Technology-OMRON)数据集上,F-measure指标相比于排名第2的显著性检测模型提高了1.9%,MAE(mean absolute error)指标降低了1.9%。同时,网络具有不错的速度表现,达到实时效果。结论 本文提出的显著性检测模型能够准确地分割出显著目标区域,并提供清晰的局部细节。     

融合分割先验的多图像目标语义分割

目的 在序列图像或多视角图像的目标分割中，传统的协同分割算法对复杂的多图像分割鲁棒性不强，而现有的深度学习算法在前景和背景存在较大歧义时容易导致目标分割错误和分割不一致。为此，提出一种基于深度特征的融合分割先验的多图像分割算法。方法 首先，为了使模型更好地学习复杂场景下多视角图像的细节特征，通过融合浅层网络高分辨率的细节特征来改进PSPNet-50网络模型，减小随着网络的加深导致空间信息的丢失对分割边缘细节的影响。然后通过交互分割算法获取一至两幅图像的分割先验，将少量分割先验融合到新的模型中，通过网络的再学习来解决前景/背景的分割歧义以及多图像的分割一致性。最后通过构建全连接条件随机场模型，将深度卷积神经网络的识别能力和全连接条件随机场优化的定位精度耦合在一起，更好地处理边界定位问题。结果 本文采用公共数据集的多图像集进行了分割测试。实验结果表明本文算法不但可以更好地分割出经过大量数据预训练过的目标类，而且对于没有预训练过的目标类，也能有效避免歧义的区域分割。本文算法不论是对前景与背景区别明显的较简单图像集，还是对前景与背景颜色相似的较复杂图像集，平均像素准确度（PA）和交并比（IOU）均大于95%。结论 本文算法对各种场景的多图像分割都具有较强的鲁棒性，同时通过融入少量先验，使模型更有效地区分目标与背景，获得了分割目标的一致性。     

融合区域和边缘特征的水平集水下图像分割

目的 水平集模型是图像分割中的一种先进方法,在陆地环境图像分割中展现出较好效果。特征融合策略被广泛引入到该模型框架,以拉伸目标-背景对比度,进而提高对高噪声、杂乱纹理等多类复杂图像的处理性能。然而,在水下环境中,由于水体高散射、强衰减等多因素的共同作用,使得现有图像特征及水平集模型难以适用于对水下图像的分割任务,分割结果与目标形态间存在较大差异。鉴于此,提出一种适用于水下图像分割的区域-边缘水平集模型,以提高水下图像目标分割的准确性。方法 综合应用图像的区域特征及边缘特征对水下目标进行辨识。对于区域特征,引入水下图像显著性特征;对于边缘特征,创新性地提出了一种基于深度信息的边缘特征提取方法。所提方法在融合区域级和边缘级特征的基础上,引入距离正则项对水平集函数进行规范,以增强水平集函数演化的稳定性。结果 基于YouTube和Bubblevision的水下数据集的实验结果表明,所提方法不仅对高散射强衰减的低对比度水下图像实现较好的分割效果,同时对处理强背景噪声图像也有较好的鲁棒性,与水平集分割方法（local pre-fitting,LPF）相比,分割精确度至少提高11.5%,与显著性检测方法（hierarchical co-salient detection via color names,HCN）相比,精确度提高6.7%左右。结论 实验表明区域-边缘特征融合以及其基础上的水平集模型能够较好地克服水下图像分割中的部分难点,所提方法能够较好分割水下目标区域并拟合目标轮廓,与现有方法对比获得了较好的分割结果。     

基于图割与改进水平集的目标提取方法

在Li模型的基础上引入C-V模型外部能量项重新构造能量函数,给出一种结合区域与边缘信息的变分水平集模型,结合基于图割理论的GCBAC算法,提出一种图割与改进变分水平集结合的目标提取方法。该方法能够让2种模型有机结合达到优势互补的效果。实验结果表明,该方法具有快速、鲁棒、抗噪性强等优点。     

多源域混淆的双流深度迁移学习

目的 针对深度学习严重依赖大样本的问题,提出多源域混淆的双流深度迁移学习方法,提升了传统深度迁移学习中迁移特征的适用性。方法 采用多源域的迁移策略,增大源域对目标域迁移特征的覆盖率。提出两阶段适配学习的方法,获得域不变的深层特征表示和域间分类器相似的识别结果,将自然光图像2维特征和深度图像3维特征进行融合,提高小样本数据特征维度的同时抑制了复杂背景对目标识别的干扰。此外,为改善小样本机器学习中分类器的识别性能,在传统的softmax损失中引入中心损失,增强分类损失函数的惩罚监督能力。结果 在公开的少量手势样本数据集上进行对比实验,结果表明,相对于传统的识别模型和迁移模型,基于本文模型进行识别准确率更高,在以DenseNet-169为预训练网络的模型中,识别率达到了97.17%。结论 利用多源域数据集、两阶段适配学习、双流卷积融合以及复合损失函数,构建了多源域混淆的双流深度迁移学习模型。所提模型可增大源域和目标域的数据分布匹配率、丰富目标样本特征维度、提升损失函数的监督性能,改进任意小样本场景迁移特征的适用性。