A Lightweight Convolutional Neural Network with Hierarchical Multi-Scale Feature Fusion for Image Classification

Convolutional neural networks (CNNs) are widely used in image classification tasks, but their increasing model size and computation make them challenging to implement on embedded systems with constrained hardware resources. To address this issue, the MobileNetV1 network was developed, which employs depthwise convolution to reduce network complexity. MobileNetV1 employs a stride of 2 in several convolutional layers to decrease the spatial resolution of feature maps, thereby lowering computational costs. However, this stride setting can lead to a loss of spatial information, particularly affecting the detection and representation of smaller objects or finer details in images. To maintain the trade-off between complexity and model performance, a lightweight convolutional neural network with hierarchical multi-scale feature fusion based on the MobileNetV1 network is proposed. The network consists of two main subnetworks. The first subnetwork uses a depthwise dilated separable convolution (DDSC) layer to learn imaging features with fewer parameters, which results in a lightweight and computationally inexpensive network. Furthermore, depthwise dilated convolution in DDSC layer effectively expands the field of view of filters, allowing them to incorporate a larger context. The second subnetwork is a hierarchical multi-scale feature fusion (HMFF) module that uses parallel multi-resolution branches architecture to process the input feature map in order to extract the multi-scale feature information of the input image. Experimental results on the CIFAR-10, Malaria, and KvasirV1 datasets demonstrate that the proposed method is efficient, reducing the network parameters and computational cost by 65.02% and 39.78%, respectively, while maintaining the network performance compared to the MobileNetV1 baseline.

改进Multi-scale ResNet的蔬菜叶部病害识别

基于深度网络的蔬菜叶部病害图像识别模型虽然性能显著,但由于存在参数量巨大、训练时间长、存储成本与计算成本过高等问题,仍然难以部署到农业物联网的边缘计算设备、嵌入式设备、移动设备等硬件资源受限的领域。该研究在残差网络(ResNet18)的基础上,提出了改进型的多尺度残差(Multi-scale ResNet)轻量级病害识别模型,通过增加多尺度特征提取模块,改变残差层连接方式,将大卷积核分解,进行群卷积操作,显著减少了模型参数、降低了存储空间和运算开销。结果表明,在PlantVillage和AI Challenge2018中15种病害图像数据集中取得了95.95%的准确率,在自采集的7种真实环境病害图像数据中取得了93.05%的准确率,在准确率较ResNet18下降约3.72%的情况下,模型的训练参数减少93%左右,模型总体尺寸缩减约35%。该研究提出的改进型Multi-scale ResNet使蔬菜叶部病害识别模型具备了在硬件受限的场景下部署和运行的能力,平衡了模型的复杂度和识别精度,为基于深度网络模型的病害识别系统进行边缘部署提供了思路。

基于多尺度Scale-Unet的单样本图像翻译

随着生成对抗网络(GAN)的发展,基于单样本的无监督图像到图像翻译(UI2I)取得了重大进展。然而,以前方法无法捕获图像中的复杂纹理并保留原始内容信息。为解决这个问题,提出了一种基于尺度可变U-Net结构(Scale—Unet)的新型单样本图像翻译结构SUGAN。所提出的SUGAN使用Scale—Unet作为生成器,利用多尺度结构和渐进方法不断改进网络结构,以从粗到细地学习图像特征。同时,提出了尺度像素损失scale-pixel来更好地约束保留原始内容信息,防止信息丢失。实验表明,与SinGAN、TuiGAN、TSIT、StyTR2等公共数据集Summer■Winter、Horse■Zebra上的方法相比,该方法生成图像的SIFID值平均降低了30%。所提方法可更好地保留图像内容信息,同时生成详细逼真的高质量图像。

融合Multi-scale CNN和Bi-LSTM的人脸表情识别研究

为了有效改善现有人脸表情识别模型中存在信息丢失严重、特征信息之间联系不密切的问题,提出一种融合多尺度卷积神经网络(Multi-scale CNN)和双向长短期记忆(Bi-LSTM)的模型。Bi-LSTM可以增强特征信息间的联系与信息的维持,在Multi-scale CNN中通过不同尺度的卷积核可以提取到更加丰富的特征信息,并通过加入批标准化(BN)层与特征融合处理,从而加快网络的收敛速度,有利于特征信息的重利用,再将两者提取到的特征信息进行融合,最后将改进的正则化方法应用到目标函数中,减小网络复杂度和过拟合。在JAFFE和FER-2013公开数据集上进行实验,准确率分别达到了95.455%和74.115%,由此证明所提算法的有效性和先进性。

基于改进Multi-Scale AlexNet的番茄叶部病害图像识别

番茄同种病害在不同发病阶段表征差异明显,不同病害又表现出一定的相似性,传统模式识别方法不能体现病害病理表征的动态变化,实用性较差。针对该问题,基于卷积神经网络提出一种适用于移动平台的多尺度识别模型,并基于此模型开发了面向农业生产人员的番茄叶部病害图像识别系统。该文详细描述了AlexNet的结构,分析其不足,结合番茄病害叶片图像特点,去除局部响应归一化层、修改全连接层、设置不同尺度卷积核提取特征,设计了基于AlexNet的多感受野识别模型,并基于Android实现了使用此模型的番茄叶部病害图像识别系统。Multi-ScaleAlexNet模型运行所耗内存为29.9MB,比原始AlexNet的内存需求652MB降低了95.4%,该模型对番茄叶部病害及每种病害早中晚期的平均识别准确率达到92.7%,基于此模型的Andriod端识别系统在田间的识别率达到89.2%,能够满足生产实践中移动平台下的病害图像识别需求。研究结果可为基于卷积神经网络的作物病害图像识别提供参考,为作物病害的自动化识别和工程化应用参考。

The Multi-Scale Numerical Modeling System for Research on the Relationship between Urban Planning and Meteorological Environment

Considering the urban characteristics, a customized multi-scale numerical modeling system is established to simulate the urban meteorological environment. The system mainly involves three spatial scales: the urban scale, urban sub-domain scale, and single to few buildings scale. In it, different underlying surface types are employed, the building drag factor is used to replace its roughness in the influence on the urban wind field, the effects of building distribution, azimuth and screening of shortwave radiation are added, and the influence of anthropogenic heating is also taken into account. All the numerical tests indicate that the simulated results are reasonably in agreement with the observational data, so the system can be used to simulate the urban meteorological environment. Making use of it, the characteristics of the meteorological environment from the urban to urban sub-domain scales, even the among-buildings scale, can be recognized. As long as the urban planning scheme is given, the corresponding simulated results can be obtained so as to meet the need of optimizing urban planning.

Study on Multi-Scale Blending Initial Condition Perturbations for a Regional Ensemble Prediction System

An initial conditions （ICs） perturbation method was developed with the aim to improve an operational regional ensemble prediction system （REPS）. Three issues were identified and investigated： （1） the impacts of perturbation scale on the ensemble spread and forecast skill of the REPS; （2） the scale characteristic of the IC perturbations of the REPS; and （3） whether the REPS＇s skill could be improved by adding large-scale information to the IC perturbations. Numerical experiments were conducted to reveal the impact of perturbation scale on the ensemble spread and forecast skill. The scales of IC perturbations from the REPS and an operational global ensemble prediction system （GEPS） were analyzed. A ＂multi-scale blending＂ （MSB） IC perturbation scheme was developed, and the main findings can be summarized as follows： The growth rates of the ensemble spread of the REPS are sensitive to the scale of the IC perturbations; the ensemble forecast skills can benefit from large-scale perturbations; the global ensemble IC perturbations exhibit more power at larger scales, while the regional ensemble IC perturbations contain more power at smaller scales; the MSB method can generate IC perturbations by combining the small-scale component from the REPS and the large-scale component from the GEPS; the energy norm growth of the MSB-generated perturbations can be appropriate at all forecast lead times; and the MSB-based REPS shows higher skill than the original system, as determined by ensemble forecast verification.

Gearbox Fault Diagnosis using Adaptive Zero Phase Time-varying Filter Based on Multi-scale Chirplet Sparse Signal Decomposition

When used for separating multi-component non-stationary signals, the adaptive time-varying filter(ATF) based on multi-scale chirplet sparse signal decomposition(MCSSD) generates phase shift and signal distortion. To overcome this drawback, the zero phase filter is introduced to the mentioned filter, and a fault diagnosis method for speed-changing gearbox is proposed. Firstly, the gear meshing frequency of each gearbox is estimated by chirplet path pursuit. Then, according to the estimated gear meshing frequencies, an adaptive zero phase time-varying filter(AZPTF) is designed to filter the original signal. Finally, the basis for fault diagnosis is acquired by the envelope order analysis to the filtered signal. The signal consisting of two time-varying amplitude modulation and frequency modulation(AM-FM) signals is respectively analyzed by ATF and AZPTF based on MCSSD. The simulation results show the variances between the original signals and the filtered signals yielded by AZPTF based on MCSSD are 13.67 and 41.14, which are far less than variances (323.45 and 482.86) between the original signals and the filtered signals obtained by ATF based on MCSSD. The experiment results on the vibration signals of gearboxes indicate that the vibration signals of the two speed-changing gearboxes installed on one foundation bed can be separated by AZPTF effectively. Based on the demodulation information of the vibration signal of each gearbox, the fault diagnosis can be implemented. Both simulation and experiment examples prove that the proposed filter can extract a mono-component time-varying AM-FM signal from the multi-component time-varying AM-FM signal without distortion.

Multi-scale Cyclone Activity in the Changjiang River–Huaihe River Valleys during Spring and Its Relationship with Rainfall Anomalies

Based on the recognition framework of the outermost closed contours of cyclones, an automated identification algorithm capable of identifying the multi-scale cyclones that occur during spring in the Changjiang River-Huaihe River valleys （CHV） were developed. We studied the characteristics of the multi-scale cyclone activity that affects CHV and its relationship with rainfall during spring since 1979. The results indicated that the automated identification algorithm for cyclones proposed in this paper could intuitively identify multi-scale cyclones that affect CHV. The algorithm allows for effectively describing the shape and coverage area of the closed contours around the periphery of cyclones. We found that, compared to the meso- and sub-synoptic scale cyclone activities, the synoptic-scale cyclone activity showed more intimate correlation with the overall activity intensity of multi-scale CHV cyclones during spring. However, the frequency of occurrence of sub-synoptic scale cyclones was the highest, and their effect on changes in CHV cyclone activity could not be ignored. Based on the area of impact and the depth of the cyclones, the sub-synoptic scale, synoptic scale and comprehensive cyclone intensity indices were further defined, which showed a positive correlation with rainfall in CHV during spring. Additionally, the comprehensive cyclone intensity index was a good indicator of strong rainfall events.

Application of Multi-Scale Tracking Radar Echoes Scheme in Quantitative Precipitation Nowcasting

A new radar echo tracking algorithm known as multi-scale tracking radar echoes by cross-correlation （MTREC） was developed in this study to analyze movements of radar echoes at different spatial scales. Movement of radar echoes, particularly associated with convective storms, exhibits different characteristics at various spatial scales as a result of complex interactions among meteorological systems leading to the formation of convective storms. For the null echo region, the usual correlation technique produces zero or a very small magnitude of motion vectors. To mitigate these constraints, MTREC uses the tracking radar echoes by correlation （TREC） technique with a large ＂box＂ to determine the systematic movement driven by steering wind, and MTREC applies the TREC technique with a small ＂box＂ to estimate small-scale internal motion vectors. Eventually, the MTREC vectors are obtained by synthesizing the systematic motion and the small-scale internal motion. Performance of the MTREC technique was compared with TREC technique using case studies： the Khanun typhoon on 11 September 2005 observed by Wenzhou radar and a squall-line system on 23 June 2011 detected by Beijing radar. The results demonstrate that more spatially smoothed and continuous vector fields can be generated by the MTREC technique, which leads to improvements in tracking the entire radar reflectivity pattern. The new multi-scMe tracking scheme was applied to study its impact on the performance of quantitative precipitation nowcasting. The location and intensity of heavy precipitation at a 1-h lead time was more consistent with quantitative precipitation estimates using radar and rain gauges.

A Multi-Scale Urban Atmospheric Dispersion Model for Emergency Management

To assist emergency management planning and prevention in case of hazardous chemical release into the atmosphere,especially in densely built-up regions with large populations,a multi-scale urban atmospheric dispersion model was established.Three numerical dispersion experiments,at horizontal resolutions of 10 m,50 m and 3000 m,were performed to estimate the adverse effects of toxic chemical release in densely built-up areas.The multi-scale atmospheric dispersion model is composed of the Weather Forecasting and Research （WRF） model,the Open Source Field Operation and Manipulation software package,and a Lagrangian dispersion model.Quantification of the adverse health effects of these chemical release events are given by referring to the U.S.Environmental Protection Agency＇s Acute Exposure Guideline Levels.The wind fields of the urban-scale case,with 3 km horizontal resolution,were simulated by the Beijing Rapid Update Cycle system,which were utilized by the WRF model.The sub-domain-scale cases took advantage of the computational fluid dynamics method to explicitly consider the effects of buildings.It was found that the multi-scale atmospheric dispersion model is capable of simulating the flow pattern and concentration distribution on different scales,ranging from several meters to kilometers,and can therefore be used to improve the planning of prevention and response programs.

An Application of the RAMS/FLUENT System on the Multi-Scale Numerical Simulation of the Urban Surface Layer—A Preliminary Study

The Regional Atmospheric Modeling System （RAMS） and the computational fluid dynamics （CFD） codes known as FLUENT are combinatorially applied in a multi-scale numerical simulation of the urban surface layer （USL）. RAMS and FLUENT are combined as a multi-scale numerical modeling system, in which the RAMS simulated data are delivered to the computational model for FLUENT simulation in an offline way. Numerical simulations are performed to present and preliminarily validate the capability of the multi-scale modeling system, and the results show that the modeling system can reasonably provide information on the meteorological elements in an urban area from the urban scale to the city-block scale, especially the details of the turbulent flows within the USL.

Bayesian seismic multi-scale inversion in complex Laplace mixed domains

Seismic inversion performed in the time or frequency domain cannot always recover the long-wavelength background of subsurface parameters due to the lack of low-frequency seismic records. Since the low-frequency response becomes much richer in the Laplace mixed domains, one novel Bayesian impedance inversion approach in the complex Laplace mixed domains is established in this study to solve the model dependency problem. The derivation of a Laplace mixed-domain formula of the Robinson convolution is the first step in our work. With this formula, the Laplace seismic spectrum, the wavelet spectrum and time-domain reflectivity are joined together. Next, to improve inversion stability, the object inversion function accompanied by the initial constraint of the linear increment model is launched under a Bayesian framework. The likelihood function and prior probability distribution can be combined together by Bayesian formula to calculate the posterior probability distribution of subsurface parameters. By achieving the optimal solution corresponding to maximum posterior probability distribution, the low-frequency background of subsurface parameters can be obtained successfully. Then, with the regularization constraint of estimated low frequency in the Laplace mixed domains, multi-scale Bayesian inversion inthe pure frequency domain is exploited to obtain the absolute model parameters. The effectiveness, anti-noise capability and lateral continuity of Laplace mixed-domain inversion are illustrated by synthetic tests. Furthermore,one field case in the east of China is discussed carefully with different input frequency components and different inversion algorithms. This provides adequate proof to illustrate the reliability improvement in low-frequency estimation and resolution enhancement of subsurface parameters, in comparison with conventional Bayesian inversion in the frequency domain.

Experimental study on spectrum and multi-scale nature of wall pressure and velocity in turbulent boundary layer

When using a miniature single sensor boundary layer probe, the time sequences of the stream-wise velocity in the turbulent boundary layer （TBL） are measured by using a hot wire anemometer. Beneath the fully developed TBL, the wall pressure fluctuations are attained by a microphone mechanism with high spatial resolution. Analysis on the statistic and spectrum properties of velocity and wall pressure reveals the relationship between the wall pressure fluctuation and the energy-containing structure in the buffer layer of the TBL. Wavelet transform shows the multi-scale natures of coherent structures contained in both signals of velocity and pressure. The most intermittent wall pressure scale is associated with the coherent structure in the buffer layer. Meanwhile the most energetic scale of velocity fluctuation at y＋ = 14 provides a specific frequency f9 ≈ 147 Hz for wall actuating control with Ret = 996.

SIMULATION OF CRACK DIAGNOSIS OF ROTOR BASED ON MULTI-SCALE SINGUUR-SPECTRUM ANALYSIS

In the diagnosis of rotor crack based on wavelet analysis, it is a painful task to find out an adaptive mother wavelet as many of them can be chosen and the analytic results of different mother wavelets are yet not the same. For this limitation of wavelet analysis, a novel diagnostic approach of rotor crack based on multi-scale singular-spectrum analysis （MS-SSA） is proposed. Firstly, a Jeffcott model of a cracked rotor is developed and the forth-order Runge-Kutta method is used to solve the motion equations of this rotor to obtain its time response （signals）. Secondly, a comparatively simple approach of MS-SSA is presented and the empirical orthogonal functions of different orders in various scales are regarded as analyzing functions. At last, the signals of the cracked rotor and an uncracked rotor are analyzed using the proposed approach of MS-SSA, and the simulative results are compared. The results show that, the data-adaptive analyzing functions can capture many features of signals and the rotor crack can be identified and diagnosed effectively by comparing the analytic results of signals of the cracked rotor with those of the uncracked rotor using the analyzing functions of different orders.

Denoising of seismic data via multi-scale ridgelet transform

Noise has traditionally been suppressed or eliminated in seismic data sets by the use of Fourier filters and, to a lesser degree, nonlinear statistical filters. Although these methods are quite useful under specific conditions, they may produce undesirable effects for the low signal to noise ratio data. In this paper, a new method, multi-scale ridgelet transform, is used in the light of the theory of ridgelet transform. We employ wavelet transform to do sub-band decomposition for the signals and then use non-linear thresholding in ridgelet domain for every block. In other words, it is based on the idea of partition, at sufficiently fine scale, a curving singularity looks straight, and so ridgelet transform can work well in such cases. Applications on both synthetic data and actual seismic data from Sichuan basin, South China, show that the new method eliminates the noise portion of the signal more efficiently and retains a greater amount of geologic data than other methods, the quality and consecutiveness of seismic event are improved obviously as well as the quality of section is improved.

Multi-scale variability of subsurface temperature in the South China Sea

Using Morlet wavelet transform and harmonic analysis the multi-scale variability of subsurface temperature in the South China Sea is studied by analyzing one-year (from April 1998 to April 1999) ATLAS mooring data. By wavelet transform, annual and semi-annual cycle as well as intrasea-sonal variations are found, with different dominance, in subsurface temperature. For annual harmonic cycle, both the downward net surface heat flux and thermocline vertical movement partially control the subsurface temperature variability. For semi-annual cycle and intraseasonal variability, the subsurface temperature variability is mainly linked to the vertical displacement of thermocline.

Multi-scale Fractal Characteristics of Atmospheric Boundary-Layer Turbulence

The turbulence data are decomposed to multi-scales and its respective fractal dimensions are computed. The conclusions are drawn from investigating the variation of fractal dimensions. With the level of decomposition increasing, the low-frequency part extracted from the turbulence signals tends to be simple and smooth, the dimensions decrease; the high-frequency part shows complex, the dimensions are fixed, about 1.70 on the average, which indicates clear self-similarity characteristics.

Multi-scale modeling of hemodynamics in the cardiovascular system

The human cardiovascular system is a closed- loop and complex vascular network with multi-scaled het- erogeneous hemodynamic phenomena. Here, we give a selective review of recent progress in macro-hemodynamic modeling, with a focus on geometrical multi-scale model- ing of the vascular network, micro-hemodynamic modeling of microcirculation, as well as blood cellular, subcellular, endothelial biomechanics, and their interaction with arter- ial vessel mechanics. We describe in detail the methodology of hemodynamic modeling and its potential applications in cardiovascular research and clinical practice. In addition, we present major topics for future study： recent progress of patient-specific hemodynamic modeling in clinical applica- tions, micro-hemodynamic modeling in capillaries and blood cells, and the importance and potential of the multi-scale hemodynarnic modeling.

Surrogate-based modeling and dimension reduction techniques for multi-scale mechanics problems

Successful modeling and/or design of engineering systems often requires one to address the impact of multiple ＂design variables＂ on the prescribed outcome.There are often multiple,competing objectives based on which we assess the outcome of optimization.Since accurate,high fidelity models are typically time consuming and computationally expensive,comprehensive evaluations can be conducted only if an efficient framework is available.Furthermore,informed decisions of the model/hardware＇s overall performance rely on an adequate understanding of the global,not local,sensitivity of the individual design variables on the objectives.The surrogate-based approach,which involves approximating the objectives as continuous functions of design variables from limited data,offers a rational framework to reduce the number of important input variables,i.e.,the dimension of a design or modeling space.In this paper,we review the fundamental issues that arise in surrogate-based analysis and optimization,highlighting concepts,methods,techniques,as well as modeling implications for mechanics problems.To aid the discussions of the issues involved,we summarize recent efforts in investigating cryogenic cavitating flows,active flow control based on dielectric barrier discharge concepts,and lithium（Li）-ion batteries.It is also stressed that many multi-scale mechanics problems can naturally benefit from the surrogate approach for ＂scale bridging.＂