Blood cell counting based on U-Net++ and YOLOv5

Clinical information about a variety of disorders is available through blood cell counting, which is usually done by manual methods. However, manual methods are complex, time-consuming and susceptible to the subjective experience of inspectors. Although many efforts have been made to develop automated blood cell counting algorithms, the complexity of blood cell distribution and the highly overlapping nature of some red blood cells (RBCs) remain significant challenges that limit the improvement of analytical accuracy. Here, we proposed an end-to-end method for blood cell counting based on deep learning. Firstly, U-Net++ was used to segment the whole blood cell image into several regions of interest (ROI), and each ROI contains only one single cell or multiple overlapping cells. Subsequently, YOLOv5 was used to detect blood cells in each ROI. Specifically, we proposed several strategies, including fine classification of RBCs, adaptive adjustment for non-maximal suppression (NMS) threshold and blood cell morphology constraints to improve the accuracy of detection. Finally, the detection outcomes for each ROI were combined and superimposed. The results show that our method can effectively address the issue of high overlap and precisely segment and detect blood cells, with a 98.18% accuracy rate for blood cell counting.     

Doped/Undoped A1-A2 Typed Copolymers as ETLs for Highly Efficient Organic Solar Cells

The electron transport layer (ETL) is a critical component in achieving high device performance and stability in organic solar cells. Conjugated polyelectrolytes (CPEs) have become an attractive alternative due to film-forming properties and ease of preparation. However, p-type CPEs generally exhibit poor charge mobility and conductivity, incorporation of electron-withdrawing units forming alternated D-A conjugated backbone can make up for these deficiencies. Herein, the ratio of electron withdrawing moieties are further increased and two poly(A1-alt-A2) typed PIIDNDI-Br and PDPPNDI-Br based on the combination of naphthalene diimide (NDI) with isoindigo (IID) or diketopyrrolopyrrole (DPP) via direct arylation polycondensation are synthesized. These CPEs possess excellent alcohol solubility, a suitable lowest unocuppied molecular orbital energy level, and work function tunability. Surprisingly, the incorporation of IID and DPP units generate distinct self-doping behaviors, which are confirmed by UV–vis absorption and ESR spectra. However, no matter doped or undoped, both CPEs present better charge-transporting properties and conductivity when utilized as ETLs. The PIIDNDI-Br and PDPPNDI-Br display good universal compatibility with the blend of PM6:Y6 and PM6:L8-BO, and PCEs of 18.32% and 18.36% are obtained, respectively, which also present excellent storage stability. In short, the combination of two different acceptors demonstrates an efficient strategy to design highly efficient ETLs for high performance photovoltaic devices.     

Tailoring of Photoluminescence Properties in All-Vacuum Deposited Perovskite via Ruddlesden–Popper Faults

Ruddlesden–Popper (RP) faults are well known in oxide perovskites, and are also observed in promising metal halide perovskites. However, the effect of RP faults on optical properties of perovskite has not been systematically investigated. In this study, it is found that RP faults are common planar faults in all-vacuum deposited CsPbBr₃-based perovskite polycrystal thin films, and the density of RP planar faults can be greatly increased by non-stoichiometric composition (Cs-rich) as well as reduced dimensionality (quasi-2D) strategies. The photoluminescence (PL) measurement reveals monotonically increasing peak intensities with higher densities of RP planar faults from Cs-rich, quasi-2D to Cs-rich & quasi-2D samples. The corresponding atomic-scale differential phase contrast maps indicate strongly confined charges within the RP planar fault network, which explains well the relationship between PL enhancement and the density of RP planar faults, and offers an alternative pathway for tailoring the optoelectronic properties of perovskite.     

双循环策略下岸桥与跨运车的联合调度

集装箱码头采用跨运车能够减少作业环节和码头机械设备的种类与数量,同时缓存区容量的设置至关重要。首先,为降低码头总完工时间、提高码头作业效率,并解决采用跨运车作为水平运输设备与岸桥进行联合装卸作业时产生的时空协调问题,引入了双循环操作策略,对岸桥与跨运车的联合作业序列优化问题进行研究。其次,建立以总完工时间最小化为目标,考虑岸桥与跨运车双循环操作的实际约束、岸桥缓存区容量限制、安全时间等约束的混合整数规划模型。然后,针对传统禁忌搜索（TS）算法的局限性,加入贪婪算法、多种邻域搜索方式、响应性策略,设计了基于贪婪算法的响应性TS算法,并进行了数值实验。实验结果验证了所提模型与算法的有效性。最后,通过对缓存区容量与跨运车数量、岸桥与跨运车配比的实验分析,得出了最优的跨运车数量和缓存区容量、岸桥与跨运车配比。结果表明：与传统码头设备配置相比,双循环策略可减少跨运车使用数量,提高岸桥与跨运车使用率。     

基于深度神经网络的多指灵巧手抓取手势优化

基于深度神经网络模型,提出了一种适用于多指灵巧手的抓取手势优化方法。首先,在仿真环境下构建了一个抓取数据集,并在此基础上训练了一个卷积神经网络,依据目标物体单目视觉信息和多指灵巧手抓取位形来预测抓取质量函数,由此可以将多指灵巧手的抓取规划问题转化为使抓取质量最大化的优化问题,进一步,基于深度神经网络中的反向传播和梯度上升算法实现多指灵巧手抓取手势的迭代与优化。在仿真环境中,比较该网络和仿真平台对同一抓取位形的抓取质量评估结果,再利用所提出的优化方法对随机搜索到的初始手势进行优化,比较优化前后手势的力封闭指标。最后,在实际机器人平台上验证本文方法的优化效果,结果表明,本文方法对未知物体的抓取成功率在80%以上,对于失败的抓取,优化后成功的比例达到90%。     

面向样本扩充的新型风格迁移网络研究

全监督语义分割网络在训练时需要耗费大量的人力与时间成本来标注样本。所以减少人工标注样本的时间,同时提升语义分割效果,对于深度学习网络的快速部署和应用推广具有重要意义。提出一种基于改进图像风格迁移网络（CycleGAN-AD）的样本扩充方法。以CycleGAN为基础,在生成器中引入注意力机制并将深度残差网络改为密集连接卷积网络。利用计算机批量产生自带标签的模拟样本,使用CycleGAN-AD网络将模拟样本风格迁移成为真实样本风格（标签不变）,并用于扩充训练样本。对石墨电极的钢印字符进行语义分割的实验结果表明,采用CycleGAN-AD网络进行样本扩充后,其分割效果得到显著提升,MIoU值最高升至0.826 0。可见,提出的样本扩充方法有希望在显著减少人工标注工作量的同时,获得高质量的训练样本。     

基于CNN-LSTM双流融合网络的危险行为识别

针对目前人体危险行为识别过程中由于时空特征挖掘不充分导致精度不够的问题,对传统双流卷积模型进行改进,提出了一种基于CNN-LSTM的双流卷积危险行为识别模型。该模型将CNN网络与LSTM网络并联,其中CNN网络作为空间流,将人体骨架空间运动姿态分为静态与动态特征进行分别提取,两者融合作为空间流的输出;在时间流中采用改进的可滑动长短时记忆网络,以增加人体骨架时序特征的提取能力;最后将两个分支进行时空融合,利用Softmax对危险动作做出分类识别。在公开的NTU-RGB+D数据集和Kinetics数据集上的实验结果表明,改进后模型的平均跨角度（Cross view,CV）精度达到92.5%,平均跨视角（Cross subject,CS）精度为87.9%。所提方法优于改进前及其他方法,可以有效地对人体危险动作做出识别,同时对于模糊动作也有较好的区分效果。     

基于自监督学习的变电站缺陷检测

为了在使用少量标注样本情况下提升变电站设备缺陷检测精度,提出一种基于自监督模型SimSiam的改进缺陷检测算法.不同于原始SimSiam,改进后的算法无需使用标志性图像(iconic images),如ImageNet数据集,而是直接利用非标志性图像(non-iconic images)如COCO数据集进行对比学习,并在下游的缺陷检测任务上获得可媲美有监督方法的性能.通过在投影层(projection head)和预测层(prediction head)中使用全卷积网络和空间注意力模块来代替MLP,保留高维特征的空间结构及局部信息;同时在计算相似度前先对特征图进行均值池化以得到特征向量,并对特征向量进行归一化以计算欧氏距离,从而改进了自监督对比学习的损失函数.实验结果表明该算法能充分利用非标志性图像进行对比学习,并在只标注少量样本的条件下提升变电站设备缺陷检测的精度,在表计表盘破损、挂空悬浮物、鸟巢、呼吸器硅胶变色及箱门闭合异常等5类缺陷检测任务上mAP达到83.84%.     

深度学习屈光检测方法研究

屈光不正是一种非常常见且对视功能发育有严重危害的眼科问题。准确与方便的屈光检测技术,对于及时发现屈光不正问题以及采取相应措施进行干预具有非常重要的意义。目前的屈光筛查设备虽然能较快进行屈光检测,但主要存在两个问题：检测准确度较低,对被测者配合度要求较高。因此,提出一种新的屈光检测方法,此方法使用基于偏心摄影验光原理的光学系统获取人脸面部近红外图像,使用图像处理技术对面部近红外图像进行处理,得到左右瞳孔图像和瞳孔位置信息,使用提出的结合了深度可分离卷积和SE模块的混合数据多输入神经网络模型进行训练与屈光度的计算。与传统偏心摄影验光原理的屈光检测方法相比,此方法有望随着数据集的扩增而达到更高的准确度,并且此方法将瞳孔位置信息作为模型的输入,可以解决传统算法对被测者配合度要求较高的问题。该研究是对屈光检测新方法的一种有益探索,使用此方法有利于屈光筛查更便利地进行,为实现非接触自助式的屈光筛查提供基础。     

基于IGWO-XGBoost融合模型的沥青路面抗滑性能评估

为研究沥青路面抗滑性能影响因素,精确预测路面抗滑性能,本文使用Gocator 3110三维智能传感器采集沥青混合料试件表面纹理并使用摆式摩擦仪测试试件表面摩擦系数.针对三维纹理点云数据中的异常数据,提出基于径向基函数(RBF)的邻域插值算法进行数据质量提升.根据修复后的三维纹理点云数据计算出具有代表性的10类宏观纹理特征参数,并采用Pearson系数相关性分析法去除冗余因子,改进模型的输入特征,并构建基于改进灰狼优化算法(IGWO)与XGBoost融合的沥青路面抗滑性能预测模型,预测沥青路面的摩擦系数.结果表明,提出模型的预测精度优于多元线性回归模型、支持向量机回归模型以及基于网格化搜索的XGBoost模型.