Auto spray painting robots can realize the automation of various types of auto lamp spraying. The auto lamp detection algorithm is the key technology of the robot. The Faster R-CNN model is selected to detect the lamp, and some improvements are made to the model to achieve better detection effect. To generate more multi-dimensional feature expressions for auto lamps, residual network ResNet101 and feature pyra-mid network (FPN) are used to extract the features of auto lamps. Compatible Intersection over Union (CIoU) is used as the bounding box regression loss function to provide moving direction and more accurate position information for the bounding box and accelerate the convergence speed of the model. Through the auto lamp detection experiment, the average recognition accuracy of the improved model is 98.96%, and the detection speed is 0.46s per image. The improved model can realize the effective recognition of auto lamp types and meet the requirements of real-time processing.
Aiming at the problems that the current RANSAC (Random Sample Consensus) algorithm has too large randomness and is susceptible to external point interference, which leads to the reduction of matching accuracy, an improved RANSAC algorithm combining feature matching confidence and grid clustering is proposed. Firstly, rough matching is carried out by the FLANN algorithm, and confidence analysis is carried out on the coarse matching point pairs, then expanding grid clustering around the high confidence point pairs. Multiple local optimal interior points are screened to optimize the global interior points and improve the matching accuracy of feature points. The experimental results show that the improved RANSAC in this paper increases the existence probability of interior points, avoids too many wrong feature matching affecting the model effect of the homography matrix, and improves the accuracy of feature matching.
In order to improve the professionalism and accuracy of machine translation in the field of notice to air navigation,a two-way machine translation model of notice to air navigation based on transfer learning and back translation is proposed to solve the problem of the parallel corpus of machine translation in the field of air notice.On the basis of training the machine translation of People's Daily,the parameters obtained from the training are transferred to the encoding end and the decoding end of the machine translation model of the partly back-translated announcements to initialize the parameters at both ends,and adjust the parameters at both ends through experiments.The two-way machine translation model of the notice of air travel is obtained.Experiments show that the introduction of the data enhancement strategy of transfer learning and partial back translation improves the robustness of the mod-el and improves the translation quality.The BLEU value of the translation evaluation index of the Chinese-to-English model has in-creased by 2.08%.The BLEU value of the translation evaluation index increased by 3.12%.
Scene texts with explicit semantic information in natural images can provide important clues to solve the corresponding computer vision problems. In the text, we usually focus on using multimodal content in the form of visual and text prompts to solve the task of fine-grained image classification and retrieval. In this paper, graph convolution network is used to perform multimodal reasoning, and the features of relationship enhancement are obtained by learning the common semantic space between salient objects and texts found in images. By obtaining a set of enhanced visual and textual functions, the proposed model is highly superior to the existing technologies in two different tasks (fine-grained classification and image retrieval in contextual texts).
To realize the automatic symptom recognition and classification of MR images and improve the accuracy and efficiency of the diagnosis of lumbar intervertebral disc herniation (LDH), a method for lumbar intervertebral disc recognition and disease classification is proposed in this paper. The method mainly includes three steps: preprocessing, target segmentation, and symptom classification. Preprocessing is performed by noise reduction and interference removal methods for blurred images. The contour poles are used to determine the four points of the tail vertebra in order to reduce the wrong segmentation of the tail vertebra. A classification method based on five judgment indicators is proposed, which effectively improves the stability of disease diagnosis. The example verifies that the algorithm can accurately complete the target segmentation and the accuracy of symptom classification reaches the standard of professional doctors, which proves that the method has good robustness.
To solve the problem of insufficient facial expression feature extraction by the VGG16 network, resulting in low recognition accuracy, an improved VGG16 network is proposed. Firstly, take the convolutional layer with the same channel in the VGG16 network as a block and divide it into five blocks.Then use the feature fusion method, the features extracted from the last three blocks of VGG16 network are fused to enhance the extraction of facial expression features. Finally, the attention mechanism SGE (spatial group enhance) module is introduced to promote the network to extract features that are conducive to facial expression recognition. At the same time, the three fully connected layers of the VGG16 network are changed to one fully connected layer to directly output the classification results, which can reduce the number of parameters while ensuring the accuracy of recognition. Experimental results show that the improved VGG16 network achieves 86.701% accuracy on RAF-DB and 56.881% accuracy on SFEW, which is higher than the original network.
With more and more applications of blockchain,how to realize interconnection between thousands of transaction blockchains,and how to realize AML,KYC and other analysis and mining has become more and more obstacles restricting the development of blockchain in the future. This paper studies a blockchain data lake architecture,through which thousands of blockchains will converge into a blockchain data lake. Through the blockchain data lake,not only the sharing of thousands of blockchains themselves can be realized. It also realizes the sharing of blockchain data and various original data of the blockchain,thus providing a feasible structure and mechanism for the analysis of the full amount of blockchain big data in the future.
区块链因其固有的防篡改、透明和可追溯等特性,被广泛地使用于对交易数据的可信管理上.目前,已经有大量的研究工作从共识算法、交易查询和溯源等方面对交易的可信管理进行了研究.为了方便后续的研究,论文对区块链实现交易数据可信管理相关的研究工作进行了梳理.首先,论文对区块链系统中确保去中心化环境下各节点间数据一致性的共识算法和便于数据管理的可信数据结构进行了分析;其次,对区块链系统上为实现交易可信查询的算法进行了梳理归纳,并对区块链在医疗数据、物联网和数据溯源上的应用分别进行了阐述;最后,针对目前区块链在交易可信管理方面的不足,分析和展望了未来的研究方向和面临的挑战.
Trustworthy transactions are the core and foundation of the modern service industry. Industries such as FinTech,digital assets trading and electronic payments are experimenting with blockchain technology to achieve trustworthiness in transactions. Blockchain as a ledger has advantages such as tamper-proof,traceability and transparency. However,the anonymity and decentralized nature of blockchain is trying to evade regulation. Therefore,the adoption of blockchain to achieve trustworthy transactions cannot be separated from the regulation of blockchain. In order to make blockchain technology better used for trusted transactions,a new type of blockchain is designed to supervise participants and transactions,and at the same time,it is able to perform transaction rollback in case of erroneous transactions and transactions found to be incompatible with the supervision conditions,and the function of ex post supervision is designed to analyze the transactions and be able to monitor the transaction status of the blockchain. Finally,the designed blockchain system is elaborated using data asset transactions as an example.
针对大部分的深度卷积神经网络存在着层次难以达到很深以及神经网络的深层次性能退化问题,论文基于前馈去噪卷积神经网络的模型,提出一种结合批量重整化去噪网络中的并行网络.新方法是将原有的网络层一分为二,通过增加网络的宽度而不是深度深层结构达到获取更多特征的目的.再使用残差学习和批量归一化的方法,以改善去噪质量,并且加速训练.结果表明,相比于当前比较成熟的去噪算法,论文提出的残差连接的并行网络去噪新算法的去噪效果更为突出,去噪耗时也大大降低.
水平井体积压裂是致密油藏高效开采的主要技术手段,准确预测产能对油田施工方案编制具有重要的指导意义,开采效果受地层因素、原油物性因素、压裂施工因素等影响,基于机理计算公式的传统预测方法存在一定局限性,提出一种基于K-means聚类与支持向量回归的产能预测组合模型,采用主成分分析算法解决K-means中欧氏距离对所有特征贡献程度一致性问题,K-means聚类结果与压裂施工参数结合作为SVR预测样本,有效解决不同区域间差异较大等问题.通过实验对比SVR、BP神经网络,预测准确性和稳定性优于单一模型,具有较高的合理性,可为致密油田高效开发提供指导性建议.
传统的主题模型在对短文本建模时,会由于词汇稀疏导致模型效果不好.论文针对短文本数据特征,提出了BERT-LDA主题挖掘模型(Short Text Topic Mining Based on BERT and LDA),该算法通过使用预训练BERT模型提取文本语义特征,再通过K-means聚类算法将短文本聚合成长文本再进行主题建模,从而扩充单条文本包含的语义特征,有效降低了词汇稀疏性,从而提升模型效果.通过在实际数据上进行对比实验证明,与LDA和BTM模型相比,该算法能够取得更低的困惑度.
目前道路目标检测算法研究多基于正常天气,对雾天场景下道路目标检测研究较少,故提出一种基于YO-LOv3的雾天道路目标检测算法.为提高检测设备在恶劣天气下对道路目标的检测能力和检测速度,利用改进后的金字塔池化结构构建去雾模块,并将其嵌入至YOLOv3目标检测网络中;引入通道注意力机制,提高DarkNet53的图像特征信息提取能力;增加一层检测层提高小目标物体检测能力,实现了一种去雾网络和检测网络联合优化的道路目标检测网络A-YO-LOv3.利用大气散射模型和图像景深信息自制雾天道路数据集S-KITTI并将其用于实验验证,实验结果表明:通过改进优化使得模型检测精度从65.22%提升至72.5%.
针对数据稀疏性问题,提出了一种新的相似度计算方法来提高传统协同过滤方法(CF)的精度.根据与用户有强相关性的用户偏好进行分析,向用户提供他们所需的项目.皮尔逊相关系数和余弦相似度,作为应用最广泛的方法,仅根据用户对项目的共同评分来发现用户之间的相关性.因此,这些方法缺乏解决稀疏性的能力.论文提出了一种新的基于全局用户偏好的相似度方法来解决稀疏性问题,提高推荐的准确性.因此,该方法的新颖之处在于能够解决相似性问题,同时能够发现不相关用户之间的关系.此外,在计算一对用户之间的相似度的过程中,为了确定正确的邻居数量,该方法考虑了两个主要因素(公平性和共评比例).并在MovieLens 100K数据集下用于评估论文算法的准确性.实验结果表明,与传统CF相似性方法相比,该方法在各项指标上都有所提高.
不同波段的红外图像既具有信息差异性也具有相似性,每个波段与不同波段的组合也包含了丰富的语义信息.因此,在目标检测时充分利用不同波段图像的信息互补性,是提高目标检测和识别能力的有效途径.论文对红外多波段图像进行像素级融合,并提出了一种多个波段并联输入、并对单波段数据使用统一模型进行数据增强的方法,对图像信息进行表征,通过构建包含了同一场景下的多个波段信息的红外图像数据集,保证了不同波段信息数据增强的一致性.论文以YOLOv4网络模型为框架,利用红外多波段数据集进行单波段图像模型与多波段图像融合模型的精度对比实验,实验结果表明,多波段数据融合算法能够有效利用其子波段图像的正向信息,相较于单个红外波段的表现,mAP提升了10%以上,验证了该方法在多波段图像目标检测与识别方面具有优势.
对风电机组运行数据中的异常数据进行检测与清洗是风电建模分析的必要前提.当原始数据中部分机组异常数据占比过大时,以往检测方法难以有效分化正常数据和异常数据,清洗过程中未能考虑运行工况,且未能输出异常关键性能参数.针对上述问题,论文提出一种基于LSTM-AE集成共享框架的风电机组异常数据检测、清洗与解释方法.该方法提出一种能在模型训练过程中优化调整各个机组数据影响比重的隐藏状态共享模块,并结合LSTM-AE网络结构设计了能有效进行多机组模型联合训练的集成共享框架,以计算异常指标重构误差;通过重构误差的在多元高斯分布中的概率密度与重构值的非线性期望函数设置自适应阈值进行异常数据清洗;对比重构误差中不同性能参数与概率密度差值的互信息量,确定异常关键性能参数.实验结果表明所提方法能提高正常数据与异常数据的分化程度,提升异常数据清洗准确率,并输出异常关键性能参数.
目前在场景文本检测领域中,基于深度学习的检测算法已经取代了传统的文本检测算法.针对在深度学习算法中基于分割的方法和基于边界框回归的两类方法被广泛的应用在文本检测当中的情形,提出了基于可变形注意力Trans-former的场景文本检测算法.首先,在采用ResNet残差网络作为骨干网络的基础上引入了Transformer编解码结构,以此将检测的目标聚焦至文本上;然后,在Transformer编解码结构中添加了可变形注意力机制,有利于让模型只关注参考点附近少量的关键采样点,降低高分辨率特征图的计算复杂度,缓解小目标文本检测困难的问题.通过在ICDAR 2013、ICDAR 2015、MSRA-TD 500与论文的场景文本数据集上的实验结果表明,与传统的深度学习主流方法相比,从F1score与Ap等实验指标上验证了算法的有效性与先进性.
为了可以快速准确地提取线结构光光条中心,提出了基于改进UNet网络的线结构光光条中心提取方法,即基于RCNN(循环卷积神经网络)单元的UNet网络模型,利用端到端的深度学习方法提取线结构光激光条纹中心.该模型将RCNN单元引入到UNet网络模型中,并代替了原来的普通CNN单元.端到端的深度学习方法避免了先分割光条后提取中心的一般过程,避免了传播错误;RCNN单元可以更好地利用空间上下文和丰富的低级视觉特征,减轻噪声对图像的影响.实验结果表明,与传统算法相比,该算法保证了光条中心的精确性和稳定性,综合性能相比传统算法明显提高.
针对传统去噪算法对边缘和纹理等细节信息保护能力不足的问题.论文提出一种基于块匹配和NLPM扩散模型的组合图像去噪算法.首先对噪声图像进行块匹配和硬阈值滤波得到基础估计图像,然后用NSST非下采样剪切波变换提取高频系数,再用NLPM模型对提取的高频系数进行扩散滤波处理,最后逆NSST重构低频子带和滤波后的高频子带得到最终去噪图像.实验表明,论文算法在峰值信噪比和结构相似度上都得到了有效的提高.
为了应对风浪、船体姿态变化等扰动因素对水炮射流造成的影响,论文使用光电成像设备、惯性测量单元和水炮等设备设计一套智能船载水炮系统,并提出一种基于射流落点与船体姿态反馈的水炮射流稳定补偿方法.一方面,建立船体姿态变化与水炮关节电机角度的补偿模型,通过惯性测量单元采集船体姿态,实现抗载体扰动;另一方面,建立射流落点与水炮关节转动角之间的逆运动学模型,根据射流落点的状态对水炮关节角进行补偿控制,实现射流的稳定射击与精准打击.仿真和实际平台实验验证表明,论文设计的系统和提出的方法能够实现射流的稳定补偿,相比于传统射流模型补偿的方法,射流对目标打击的平均误差降低了82%,提高了射流落点的准确度.
