This article studies wavelet transform-based image data fusion,select the appropriate number of decomposition layers and wavelet types.Four fusion rules were formed using two methods,weighted average and taking the maximum absolute value,to achieve pixel level image data fusion.Simulation experiments were conducted on plant images to obtain image fusion results by using MATLAB software programming.E-valuate the fusion effect of images using color features,texture features,mean,variance,and other indica-tors.The experimental results show that when using fusion rules to select the maximum absolute value of high frequency,the weighted average of low frequency,or the maximum absolute value of low frequency,the image fusion effect is better.Research based on image data fusion can obtain high-quality images,which is of great significance for improving the utilization of image information and expanding the application scenarios of images.
Empirical Mode Decomposition(EMD)is a new method of non-stationary signal processing method starting to prosper in recent years,but the end effect caused by the decomposition results in serious distortion.Therefore,an improved self-adaptive waveform matching extension method is proposed to study end effect of EMD,which can determine extension points according to different characteristics of signal ends to solve the end effect problem.Experiment results show that the improved self-adaptive waveform matching extension method can avoid the distortion of EMD of signal.The improved EMD combined with wavelet soft threshold denoising method is used to process the ultrasonic simulation signal with noise,and the signal-to-noise ratio is 33.78％higher than that of the wavelet soft threshold method,indicating that this method has certain advantages in nonlinear and non-stationary signal processing.
To solve the problems of low prediction accuracy,insufficient generalization and incomplete hy-perparameters tuning of deep learning model existed in traffic flow forecasting task,an improved ant colony algorithm based Bi-LSTM traffic flow prediction model is put forward,which uses global optimization capa-bility of the improved ant colony algorithm to optimize hyperparameters tuning towards layers of Bi-LSTM network,number of neurons,batch size,and the number of training.Experiments are carried out on two public data sets of daily traffic flow in British Motorway and Bao'an District published by Shenzhen Govern-ment Open Platform,with RMSE and MAE being as evaluation indexes.The results show that DACO-Bi-LSTM model has strong optimization ability and better prediction performance,and shows better prediction performance.
以往体温测量方式效率较低,无法满足疫情防控的需求,为了测量流动人员体温,保障公共空间的安全,提出一种基于人脸测温技术的通道闸机人员身份识别方法.利用红外热像仪采集并处理热红外人脸图像;借助人脸测温技术中的黑体模块构建温度——灰度模型,标注超温人员人脸图像;融合Gabor与SVD算法,提取超温人员人脸图像特征向量,构建与集成多分类器,识别超温人员的身份.实验数据显示,在不同特征向量维数下,应用提出方法获得的人脸测温误差较低,人员身份识别率高于最低限值,充分证实该方法的可行性.
为了解决实现云制造模式过程中的柔性作业车间调度问题,在进化算法的基础上提出了 IM-MOEA/D算法.该算法为了减少运算,种群使用双编码模式,初始化种群分两步策略和六种规则,采用两类五种变邻域搜索并设置搜索阈值,以提升算法的全局和局部迭代寻优能力.最后用算例验证了 IM-MOEA/D算法的有效性,有助于改善云制造环境下柔性作业车间调度的制造效率.
为精准获取情绪量化识别结果,研究多元线性回归与决策树的情绪量化识别算法.利用多元线性回归模型获取情绪影响因素;通过分割相似度获取最优特征,将其作为最佳分类规则,按照最佳分类规则塑造最佳决策树;在求解信息熵时,由詹森不等式替换熵函数的上凸性,剔除邻近的匹配叶节点;将情绪影响因素输入改进决策树,输出情绪量化识别结果.实验证明:该算法的平均精确匹配率高达98％,平均绝对误差低至0.18,具备较优的情绪量化识别性能.
针对现有深度学习图像去模糊方法依赖大量训练数据的问题,从图像先验信息出发,提出一种联合深度先验图像去模糊方法.用自编码-自解码网络和全连接网络,分别对潜像和模糊核的深度先验进行建模,在去模糊模型中加入TV(Total Variation)正则化和冲击滤波进行改进,从而提高方法的抗噪性能,增强复原图像细节,最后通过联合优化得到清晰图像.实验结果表明,该方法可以在不依赖大量训练数据的基础上进行图像去模糊,在公开数据集上获得的峰值信噪比值为29.24dB,结构相似度值为0.862,与其他方法相比在真实模糊图像上去模糊效果更好.
通信信号生成通常也叫信号重构,信号重构在欺骗干扰、环境构建等方面具有广泛应用,而生成对抗网络的提出为通信领域中的信号重构带来了新思路.文中利用生成对抗网络,在不对通信信号进行参数测量与特征分析的情况下,实现通信模拟调制信号的生成,并加入监督学习,使用有标签的数据集,在WGAN-GP的基础上构建CGAN与ACGAN两种网络模型进行模拟调制信号的生成,并在训练完成的生成器中通过指定标签生成特定调制样式的通信信号,最后经过对比分析,基于AC-GAN 的模拟调制信号生成的质量更优.
针对当前舰船设计企业在使用CAE计算时遇到的高性能计算资源不足的问题,设计了一种混合云架构的高性能计算平台,该平台通过整合企业内部建设的高性能计算资源和公有云资源,结合统一作业管理平台、远程可视化等技术,既能满足船舶设计企业对高性能计算资源的弹性需求,又能满足用户作业提交、作业计算进度监控和计算结果在线处理.该平台已成功应用于舰船设计企业,提高了舰船设计中CAE计算的效率,解决了高性能计算资源不足的问题.
