基于Faster R-CNN的颜色导向火焰检测

基于深度特征的目标检测方法Faster R-CNN在火焰检测任务上存在检测效率低的问题,因此提出了基于颜色引导的抛锚策略。该策略设计火焰颜色模型来限制锚的生成,即利用火焰颜色约束锚的生成区域,从而减少了初始锚的数量,提升了计算效率。为了进一步提高网络的计算效率,将区域生成网络中的卷积层替换成掩膜卷积。为了验证所提方法的检测效果,采用BoWFire和Corsician数据集进行验证。实验结果表明,该方法实际检测速度相较于原Faster R-CNN提高了10.1%,BoWFire上该方法的火焰检测F值为0.87,Corsician上该方法的准确度可达99.33%。所提方法可以提高火焰检测的效率,并能够准确检测图像中的火焰。     

Steganography using a 3-player game

Image steganography aims to securely embed secret information into cover images. Until now, adaptive embedding algorithms such as S-UNIWARD or Mi-POD, were among the most secure and most often used methods for image steganography. With the arrival of deep learning and more specifically, Generative Adversarial Networks (GAN), new steganography techniques have appeared. Among them is the 3-player game approach, where three networks compete against each other. In this paper, we propose three different architectures based on the 3-player game. The first architecture is proposed as a rigorous alternative to two recent publications. The second takes into account stego noise power. Finally, our third architecture enriches the second one with a better interaction between embedding and extracting networks. Our method achieves better results compared to existing works Hayes and Danezis (2017), Zhu et al. (2018), and paves the way for future research on this topic.     

基于深度学习的小样本中药材粉末显微图像识别

针对中药材种类繁多、数据量稀少以及难以对其导管进行分类的问题,提出一种基于多通道颜色空间与注意力机制模型的卷积神经网络改进方法。首先,采用多通道颜色空间将RGB颜色空间与其他颜色空间合并为6通道作为网络输入,使网络学习亮度、色调和饱和度等特征信息,弥补数据量的不足;其次,在网络中加入注意力机制模型,其中通道注意力模型将两个池化层紧密连接到一起,空间注意力模型将多尺度空洞卷积结合到一起,使网络将注意力聚焦于小样本中关键的特征信息。实验结果表明,针对34种中药材样本的8774张导管图像,采用多通道颜色空间和注意力机制模型的方法,与原始ResNet网络相比,准确率分别提升了1.8个百分点和3.1个百分点,将二者结合后准确率提升了4.1个百分点,说明所提方法对小样本分类的准确率有着大幅度的提升。     

Blind quality assessment for screen content images via convolutional neural network

With the wide propagation of cloud and mobile computing, screen content images (SCIs) have become more indispensable in our daily lives. Compared to natural scene images (NSIs), SCIs possess many particular characteristics, like mixed contents, extremely sharp edges, and text graphics. Consequently, more challenges occur in the feature extraction, which is used to reflect the distortion, during the quality assessment of SCIs. Recently, some convolutional neural network (CNN) models have been designed by automatically learning feature to evaluate the quality. In this paper, we develop a novel blind quality assessment method for SCIs via the CNN. First, compared with existing CNN-based methods, the proposed method avoids the disadvantage of training with image patches, and it is the pioneering attempt that takes the entire image as inputs. Second, instead of the image gray value, the original image is decomposed into two portions, i.e., the predicted and unpredicted portions, according to the internal generative mechanism (IGM) theory as the input of CNN. Through the CNN, all features of the image are learned automatically from beginning to end, and the network finally outputs the predicted score. Since existing SCI database is too small, to fully train the network, we collected 30000 SCIs and employed a high-accuracy full-reference quality assessment metric of SCI to compute scores as the training labels. Experimental results on SIQAD database demonstrate that the proposed method is comparable to reference-based SCI quality assessment metrics and is superior to the state-of-the-art NSI quality assessment metrics.     

A deep supervised approach for ischemic lesion segmentation from multimodal MRI using Fully Convolutional Network

The principle restorative step in the treatment of ischemic stroke depends on how fast the lesion is delineated from the Magnetic Resonance Imaging (MRI) images. This will serve as a vital aid to estimate the extent of damage caused to the brain cells. However, manual delineation of the lesion is time-consuming and it is subjected to intra-observer and inter-observer variability. Most of the existing methods for ischemic lesion segmentation rely on extracting handcrafted features followed by application of a machine learning algorithm. Identifying such features demand multi-domain expertise in Neuro-radiology as well as Image processing. This can be accomplished by learning the features automatically using Convolutional Neural Network (CNN). To perform segmentation, the spatial arrangement of pixel needs to be preserved in addition to learning local features of an image. Hence, a deep supervised Fully Convolutional Network (FCN) is presented in this work to segment the ischemic lesion. The highlight of this research is the application of Leaky Rectified Linear Unit activation in the last two layers of the network for a precise reconstruction of the ischemic lesion. By doing so, the network was able to learn additional features which are not considered in the existing U-Net architecture. Also, an extensive analysis was conducted in this research to select optimal hyper-parameters for training the FCN. A mean segmentation accuracy of 0.70 has been achieved from the experiments conducted on ISLES 2015 dataset. Experimental observations show that our proposed FCN method is 10% better than the existing works in terms of Dice Coefficient.     

基于双重金字塔网络的视频目标分割方法

针对复杂视频场景中难以分割特定目标的问题，提出一种基于双重金字塔网络（DPN）的视频目标分割方法。首先，通过调制网络的单向传递让分割模型适应特定目标的外观。具体而言，从给定目标的视觉和空间信息中学习一种调制器，并通过调制器调节分割网络的中间层以适应特定目标的外观变化。然后，通过基于不同区域的上下文聚合的方法，在分割网络的最后一层中聚合全局上下文信息。最后，通过横向连接的自左而右结构，在所有尺度中构建高阶语义特征图。所提出的视频目标分割方法是一个可以端到端训练的分割网络。大量实验结果表明，所提方法在DAVIS2016数据集上的性能与较先进的使用在线微调的方法相比，可达到相竞争的结果，且在DAVIS2017数据集上性能较优。     

一种基于雷达图表示的数值型数据的CNN分类方法

卷积神经网络（convolutional neural networks， CNN）是一种广泛用于分析视觉图像的分类方法.由于数值数据存在着非线性、耦合性等复杂的空间关系，因此基于CNN的数值型数据的研究较少.本文的目的是找到一种可行的方法，将CNN的应用领域扩展到数值数据.于是提出了一种基于雷达图表示的数值型数据的CNN分类方法（Radar-CNN）.该算法首先将数值数据表示成雷达图形式，然后将其输入CNN中构建分类模型.为了进一步研究特征尺度和序列对性能的影响，提出了两种改进算法Rank Radar-CNN和SFS Radar-CNN.为了验证所提算法的有效性，引入TE化工过程数据集进行实验测试并比较，实验结果表明Radar-CNN及其改进算法具有优异的性能.     

盲去模糊的多尺度编解码深度卷积网络

针对拍摄场景中物体运动不一致所带来的非均匀模糊，为提高复杂运动场景中去模糊的效果，提出一种多尺度编解码深度卷积网络。该网络采用"从粗到细"的多尺度级联结构，在模糊核未知条件下，实现盲去模糊；其中，在该网络的编解码模块中，提出一种快速多尺度残差块，使用两个感受野不同的分支增强网络对多尺度特征的适应能力；此外，在编解码之间增加跳跃连接，丰富解码端信息。与2018年国际计算机视觉与模式识别会议（CVPR）上提出的多尺度循环网络相比，峰值信噪比（PSNR）高出0.06 dB；与2017年CVPR上提出的深度多尺度卷积网络相比，峰值信噪比和平均结构相似性（MSSIM）分别提高了1.4%和3.2%。实验结果表明，该网络能快速去除图像模糊，恢复出图像原有的边缘结构和纹理细节。     

基于改进SSD的高效目标检测方法

为改善一阶段目标检测算法检测精度较差的缺陷，提出一种基于SSD的高效多目标定位检测算法FSD。该算法主要从两个方面对一阶段目标检测算法进行改进：设计了一个更高效的密集残差网络，即R-DenseNet，通过采用一种更窄的密集网络结构形式，在保持特征提取容量的同时降低了计算复杂度，从而提高了算法的检测和收敛性能；改进了损失函数，通过抑制易分样本在损失函数中的权重，提高算法的鲁棒性，改善了目标检测中样本失衡的现象。采用Tensorflow深度学习框架部署网络，并在搭载Nvidia Titan X的Ubuntu上开展实验，实验表明FSD在COCO和PASCAL VOC这两个目标检测数据集上上都取得了最高的检测精度，其中FSD300D的检测精度相比SSD300有3.7%提升，检测相率比SSD有10.87%提升。     

自然光照条件下植被几何光学四分量的提取算法

目的 几何光学四分量是指在太阳光照条件下传感器所能观测的4个光学分量，即光照植被、光照土壤、阴影植被和阴影土壤。四分量是构成遥感几何光学模型的重要内容。在近地表遥感应用中，相机俯视拍照是提取四分量的一个途径。准确快速地从图像数据中提取四分量对植被冠层结构参数反演和植被长势监测具有重要意义。方法 植被与土壤二分量的识别是四分量提取的基础。目前大多数二分类算法在自然光照条件复杂时分类误差较大。本文基于卷积神经网络（CNN）和阈值法实现了多种二分类和四分量提取算法。阈值法中，使用SHAR-LABFVC （shadow-resistant algorithm：LABFVC）实现植被与土壤的二分类，并在此基础上应用二次阈值分割获取四分量，称为二次阈值法；基于CNN的方法中，采用U-Net架构，并使用RGB和RGBV数据进行训练得到U-Net和U-Net-V模型，前者完成二分类和四分量任务，后者只完成四分量提取实验。最后，对一种结合U-Net与阈值法的混合算法进行四分量提取实验。结果 本文在18幅图像（1 800个子图）数据上进行了实验，结果表明，与目视解译得到的四分量真值相比较，U-Net-V和混合法精度最高，具有相近的均方根误差（RMSE）（0.06和0.07）和相关系数（0.95和0.94）；二次阈值法与U-Net模型精度略低于上述两种算法，RMSE分别是0.08和0.09，相关系数均为0.88。在二分类实验中，U-Net的分类正确率是91%，SHAR-LABFVC为85%。结论 通过对比实验表明，在二分类问题中，U-Net可以更好地应对复杂自然光照条件下的数字图像。在四分量提取实验中，混合法和U-Net-V的结果优于U-Net与二次阈值法，可以用于提取四分量。