基于深度卷积神经网络的视觉里程计研究

视觉里程计利用视频信息来估计相机运动的位姿参数，实现对智能体的定位。传统视觉里程计方法需要特征提取、特征匹配/跟踪、外点剔除、运动估计、优化等流程，解算非常复杂，因此，提出了基于卷积神经网络的方法来实现端对端的单目视觉里程计。借助卷积神经网络对彩色图片自动学习提取图像帧间变化的全局特征，将用于分类的卷积神经网络转化为帧间时序特征网络，通过三层全连接层输出相机的帧间相对位姿参数。在KITTI数据集上的实验结果表明，提出的Deep-CNN-VO模型可以较准确地估计车辆的运动轨迹，证明了方法的可行性。在简化了复杂模型的基础上，与传统的视觉里程计系统相比，该模型的精度也有所提高。     

On flow regime transition in trickle bed: Development of a novel deep-learning-assisted image analysis method

An image analysis method was developed based on deep-learning algorithms to extract phase fractions quantitatively in a rectangular trickle bed, and the average identification error was lower than 5%. Furthermore, the flow regime transition in the trickle bed was studied. In trickle-to-pulse flow transition, the trickle flow could be further classified into the stable trickle flow and accelerated one. The SD of liquid fractions and the peak width at half-height of the probability density curve of liquid fractions were close to zero in stable trickle flow, increased rapidly in accelerated trickle flow, and remained approximately constant in pulse flow. In bubble-to-pulse flow transition, dispersed bubbles in bubble flow induced the outliers outside the upper boundary of the boxplot of gas fraction, while alternative appearance of gas-rich zone and liquid-rich zone in pulse flow induced outliers outside both the upper and lower boundaries of the boxplot of gas fraction.     

PartitionTuner: An operator scheduler for deep-learning compilers supporting multiple heterogeneous processing units

Recently, embedded systems, such as mobile platforms, have multiple processing units that can operate in parallel, such as centralized processing units (CPUs) and neural processing units (NPUs). We can use deep-learning compilers to generate machine code optimized for these embedded systems from a deep neural network (DNN). However, the deep-learning compilers proposed so far generate codes that sequentially execute DNN operators on a single processing unit or parallel codes for graphic processing units (GPUs). In this study, we propose PartitionTuner, an operator scheduler for deep-learning compilers that supports multiple heterogeneous PUs including CPUs and NPUs. PartitionTuner can generate an operator-scheduling plan that uses all available PUs simultaneously to minimize overall DNN inference time. Operator scheduling is based on the analysis of DNN architecture and the performance profiles of individual and group operators measured on heterogeneous processing units. By the experiments for seven DNNs, PartitionTuner generates scheduling plans that perform 5.03% better than a static type-based operator-scheduling technique for SqueezeNet. In addition, PartitionTuner outperforms recent profiling-based operator-scheduling techniques for ResNet50, ResNet18, and SqueezeNet by 7.18%, 5.36%, and 2.73%, respectively.     

ETLi: Efficiently annotated traffic LiDAR dataset using incremental and suggestive annotation

Autonomous driving requires a computerized perception of the environment for safety and machine-learning evaluation. Recognizing semantic information is difficult, as the objective is to instantly recognize and distinguish items in the environment. Training a model with real-time semantic capability and high reliability requires extensive and specialized datasets. However, generalized datasets are unavailable and are typically difficult to construct for specific tasks. Hence, a light detection and ranging semantic dataset suitable for semantic simultaneous localization and mapping and specialized for autonomous driving is proposed. This dataset is provided in a form that can be easily used by users familiar with existing two-dimensional image datasets, and it contains various weather and light conditions collected from a complex and diverse practical setting. An incremental and suggestive annotation routine is proposed to improve annotation efficiency. A model is trained to simultaneously predict segmentation labels and suggest class-representative frames. Experimental results demonstrate that the proposed algorithm yields a more efficient dataset than uniformly sampled datasets.     

基于NVIDIA Jetson TX2的道路场景分割

图像语义分割是计算机视觉领域重要研究方向之一,其中基于深度学习的语义分割相较于传统分割算法更为高效可靠,可应用于交通监控、自动驾驶等领域的场景理解阶段.但复杂的分割网络在嵌入式平台上的推理速度较低,难以进行实际应用.因此针对交通监控、无人驾驶等应用背景,在嵌入式平台NVIDIA Jetson TX2上,采用基于深度卷积编解码器结构的图像分割网络,对道路场景进行语义分割,并基于NVIDIA的推理加速器TensorRT2,完成网络模型简化、网络自定义层添加与CUDA并行优化,实现了对网络推理阶段的加速.实验结果表明,加速引擎在TX2上的推理速度约为原模型的10倍,为复杂分割网络在嵌入式平台上的应用提供了支持.     

基于SVM-MLP的乳腺癌预测

乳腺癌一直是影响女性健康最重要的问题之一,已经成为全球女性发病率最高的恶性肿瘤。近年来,利用机器学习和深度学习方法来诊断癌症已经成为发展较快的一个分支。通过使用逻辑回归模型(LR)、高斯核函数支持向量机(SVM)、前馈神经网络(MLP)对同一数据集进行预测,得出其中SVM迭代时间最短,前馈神经网络预测准确率最高。为了减少前馈神经网络的迭代时间,提出了基于SVM优化的前馈神经网络分类乳腺癌模型,实验结果表明:基于SVM优化后的前馈神经网络模型与Logistic模型、传统SVM模型相比具有更高的分类准确率,且迭代时间相对减少。     

基于深度学习的气液固三相反应器图像分析方法及应用

气液固三相反应器中复杂的颗粒背景给流动参数的图像检测带来巨大挑战。发展了一种基于深度学习的气液固三相反应器图像分析方法，包括采集图像、制作训练集、建立图像识别模型和提取流动参数四个步骤。采用全卷积神经网络，在学习率为0.005、训练次数为2000次、训练集大小超过400张图像的条件下，图像识别误差小于5%。利用该方法可以获取三相反应器中局部相含率（气相分数和液相分数）及其空间分布、时间序列等信息，再采用时域分析、频率分析、小波分析等分析方法提取二次参数，可用于流型识别、压降预测和气液分布的均匀性判别等。将该方法用于涓流床中流动参数的检测，结果表明，局部液相分数的时间序列信号及其功率谱、概率密度分布均能清晰地区分涓流、脉冲流、鼓泡流等典型流型；时间序列信号的均值、标准差、极差和概率密度分布曲线半峰宽等特征参数可用于确定流型边界；平均液相分数可以用于预测涓流区的压降，计算值与实验测量值的平均相对偏差约为15%；液相分数空间分布的标准差可用于表征涓流床中不同流型的气液分布均匀性。该方法为气液固三相反应器的研究提供了新的工具。     

结合深度学习和NCFS算法的堆石料粒度分布智能检测方法

针对目前堆石坝施工过程中人工筛分试验无法实现爆堆料物粒度快速检测以及现有粒度检测模型准确度低、泛化能力差等问题,提出了一种基于深度学习模型与邻域分量特征(Neighborhood Component Feature Selection,NCFS)算法相结合的堆石坝料物粒度数字筛分检测方法,该方法可以通过拍摄料堆图像快速检测料堆粒度分布。为了提高深度学习模型的精确度,提出将基于迁移学习的Deeplabv3+模型和稠密条件随机场算法(DenseCRF)结合用于图像训练学习和优化;在料堆二维特征到三维粒度分布转换方面,提出基于NCFS算法的块石二维平面参数对三维粒度的表征公式,并采用MATLAB语言编制了相应的软件加以实现。句容抽水蓄能电站工程现场爆破料堆图像采集和筛分试验分析的结果表明:所提方法是可行的,且相比其他方法,在特征提取以及粒度检测精度上均有所提高。     

生成对抗网络辅助学习的舰船目标精细识别

针对近岸舰船目标细粒度识别的难题,提出了一种利用生成对抗网络辅助学习的任意方向细粒度舰船目标识别框架。通过训练能模仿舰船目标区域的抽象深度特征的生成网络引入生成样本,来辅助分类子网络学习样本空间的流形分布,从而增强细粒度的类别间判别能力。在细粒度类别的近岸舰船数据集上,引入生成对抗网络后的算法识别准确率得到较大提升,平均识别精度提升了2%。消融实验结果表明,利用生成样本辅助训练分类子网络可以有效地提升舰船目标的细粒度识别精度。     

基于多模态融合的三维模型检索算法研究

为了获得更好的三维模型检索分类性能,基于深度学习模型研究了多模态信息融合对三维模型的特征描述,在训练步骤提出相关性损失函数来指导不同模态之间的训练,提取更稳健的特征向量;最后将融合特征应用于三维模型的检索和分类,在ModelNet40数据集上进行了三维模型分类任务和检索任务评估。实验结果及与现有方法进行的对比证明了该方法的优越性,为三维模型检索分类领域提供了一种新的思路。