Decomposition and replacement: Spatial knowledge distillation for monocular depth estimation

Knowledge distillation has become a key technique for making smart and light-weight networks through model compression and transfer learning. Unlike previous methods that applied knowledge distillation to the classification task, we propose to exploit the decomposition-and-replacement based distillation scheme for depth estimation from a single RGB color image. To do this, Laplacian pyramid-based knowledge distillation is firstly presented in this paper. The key idea of the proposed method is to transfer the rich knowledge of the scene depth, which is well encoded through the teacher network, to the student network in a structured way by decomposing it into the global context and local details. This is fairly desirable for the student network to restore the depth layout more accurately with limited resources. Moreover, we also propose a new guidance concept for knowledge distillation, so-called ReplaceBlock, which replaces blocks randomly selected in the decoded feature of the student network with those of the teacher network. Our ReplaceBlock gives a smoothing effect in learning the feature distribution of the teacher network by considering the spatial contiguity in the feature space. This process is also helpful to clearly restore the depth layout without the significant computational cost. Based on various experimental results on benchmark datasets, the effectiveness of our distillation scheme for monocular depth estimation is demonstrated in details. The code and model are publicly available at : https://github.com/tjqansthd/Lap_Rep_KD_Depth.     

Establishment of a color tolerance for yarn-dyed fabrics from different color-depth yarns

Yarn-dyed fabric is often woven from warp and weft yarns in the same color depth to ensure a uniform color appearance. The difference in color depth between warp and weft tends to result in the uneven color of the yarn-dyed fabric. This article aims to establish a color tolerance for yarn-dyed fabric that can be woven with a qualified color appearance but from the warp and weft yarns in different color depths. A total of 27 yarn-dyed fabric samples in three color series (red, yellow, and blue) were evaluated by using the yarn-dyed fabric from warp and weft yarns in the same color depth of 2% (on weight of fabric, owf) as the standard. Visual assessment and instrumental measurement of color were carried out to establish the color tolerance ellipse that was defined as CMC (Color Measurement Committee) color differences (2:1) of no more than 1.00. It was found that the color strengths (K/S) and color differences (ΔE_CMC(2:1)) of these fabric samples for each color series had linear relationships with the color depths of warp and weft yarns. The color tolerance ellipses indicated that, even though the warp and weft yarns had an apparent color difference, they could be woven in fabrics with relatively uniform color appearance and meet the requirements for yarn-dyed fabric. This work provided valuable insight into the production of qualified yarn-dyed fabrics from unqualified dyed yarns.     

金属增材制造技术研究热点聚类分析

为深入分析金属增材制造技术分类和发展状况,采集中国期刊全文数据库(CNKI)收录的核心期刊上的768篇科技文献,借助文献分析可视化软件CiteSpace,对关键词聚类进行了全景式描绘,构建金属增材制造技术知识图谱,揭示该技术研究分类以及演化趋势。     

考虑煤层倾角和钻孔倾角的钻屑量理论研究

为准确预测与预防深部开采煤与瓦斯突出,建立了考虑煤层倾角与钻孔倾角的新钻屑量理论计算公式,并分析了煤层倾角和钻孔倾角对钻屑量的影响。以平煤股份有限公司八矿为研究对象,将钻屑量理论计算值与该矿己15-21030风巷现场钻屑量实测值进行对比验证,得出在煤层倾角一定时,钻屑量随钻孔倾角增大呈幂指数递减关系,减小的比率呈“V”形变化趋势,原钻屑量理论值与实测值平均误差为5.5%,在考虑煤层倾角与钻孔倾角情况下的新钻屑量理论值与实测值平均误差为3.9%。     

基于卷积神经网络的单目深度估计

针对利用深度学习方法对街道图像进行深度估计，提出采用语义分割的方法解决深度图出现边界模糊等问题；估计深度通过左右视角图生成视差图进行无监督的训练。在网络模型中添加语义分割层，采取多个空洞卷积并行的结构增加感受野，同时减少了图像下采样的次数，降低了由于下采样带来的信息损失，使得的结果更加准确。这也是在深度估计中首次与空洞卷积相结合增加准确率。通过对KITTI街道数据集进行训练，与现有结果相比，除了增加检测准确性，降低错误率之外，使得效果图中的物体更加清晰，并且在效果图中还保留了一些原模型中被忽视掉的细节信息，将原始图像更加完整的表现出来。     

改进的多层感知机在客户流失预测中的应用

针对在传统的客户流失预测数据预处理中，使用one-hot编码处理离散属性导致数据维度增加及数据过于稀疏的问题，提出了两种基于多层感知机的改进后的客户流失预测模型。其主要思想是分别使用堆叠自编码器和实体嵌入两种方法对多层感知机进行改进，通过将离散属性的高维编码数据向低维空间映射，有效地减少了one-hot编码产生的稀疏数据，增加了离散属性值之间的关联度。在对两份公开的数据集进行交叉验证后的实验结果表明，改进后的模型既有效地提高了预测的准确度，又维持了传统多层感知机模型在并行化计算方面的优势。     

低活化马氏体钢热变形行为及机理型本构建模研究

利用Gleeble-3500热模拟试验机研究了低活化马氏体钢在变形温度为850～950 ℃、应变速率为0.001～1 s^-1条件下的热变形行为。建立了流变应力本构方程，并评估了该方程的预测能力。绘制了低活化马氏体钢在不同应变下的热加工图。结果表明：在较高的应变速率条件下，该材料主要发生动态回复，在较高变形温度和较低应变速率下具有明显的动态再结晶特征；本构方程的预测结果与实验结果符合良好；变形温度870～930 ℃、应变速率0.001～0.01 s^-1和变形温度920～950 ℃、应变速率0.3～1 s^-1分别是真应变为0.4和0.6下最优的热加工区域。     

Propagation and termination of an internal crack in continuously cast austenitic stainless steel analyzed by ∑ values and the Schmid factor

The formation mechanism of an internal crack was clarified from the viewpoint of the crystallography and thermal expansion. An inverse pole figure map obtained by EBSD pattern showed that the crack propagated along the grain boundaries having high ∑ values within the columnar zone. After the crack initiation, these positions were considered to undergo cracking followed by propagation toward the equiaxed side. Near the termination position, the grains ahead of crack propagation had a Schmid factor higher than 0.45 consuming elastic strain energy. Thermal expansion measurements showed that the grain with (0 0 1) orientation had the largest expansion while that with (0 1 1) the smallest. The grain boundaries neighboring the combination of (0 0 1) and (0 1 1) grains had the largest thermal stress. Therefore, thermal stress contributed to the initiation of cracking. It was thus proposed to enlarge the equiaxed zone to prevent cracking by discontinuing the crack propagation.     

小尺度交通信号灯的检测与状态识别

交通信号灯识别包括检测和状态识别，在智能交通系统中发挥重要作用。基于YOLOv3算法提出了一种交通信号灯检测与状态识别模型。针对交通信号灯相较于交通场景中其他目标具有尺度小的特性进行了算法的设计：降低骨干网络的下采样倍率以增加小尺度目标的特征描述能力；通过增大特征图的尺度来改进多尺度特征融合；引入广义交并比作为检测任务的损失函数来改进目标边界框的回归效果。同时，根据交通信号灯本身的特性，使用颜色和形状约束的方法对信号灯进行状态识别和类别验证。最后在公开的Bosch交通信号灯数据集上和实际的城区道路进行了实验验证。实验结果表明，所提出的算法能够提升交通信号灯识别的精度和召回率，识别准确率可以达到90%左右。