Search

In this paper, a distributed implementation of temperature data denoising operator using the steepest descent method is presented. First, the smoothness-based denoising method using normalized Laplacian matrix is described and the conventional Neumann series implementation is reviewed briefly. Then, the steepest descent method is applied to develop a distributed implementation of denoising operator and its convergence condition is studied. It can be also shown that the Neumann series method is a special case of steepest descent method. Next, the momentum term is added to the update term of the steepest descent method for speeding up the convergence of implementation method. Finally, the real temperature data collected from the sensor network at Taiwan is used to demonstrate the effectiveness of the proposed denoising method and some discussions are made.

In this study, a Tikhonov-based graph signal denoising method is presented. First, the signal denoising problem is formulated as a minimization problem and its optimal solution is required to compute the matrix inverse which only allows the centralized processing. To avoid solving matrix inverse and obtain the distributed implementation, two methods are studied. One is the Neumann-series (NS) method; the other is the edge-variant (EV) filter. As a result, EV filter has faster convergence speed than NS method. However, at steady state, NS method has smaller approximation error than EV filter. Thus, a hybrid denoising method of NS and EV filters is proposed in this paper to get fast convergence speed and smaller approximation error simultaneously. The data from sensor network and social network are used to examine the correctness of the proposed graph signal denoising approach.

In this paper, the Katz centrality and Neumann series are used to identify the station importance of Taipei metro system. First, node importance of complex network is computed by the Katz centrality whose solution needs to solve the matrix inversion (MI). To get a MI free computation method, the truncated Neumann series expansion is then employed to approximate the MI. Next, a polynomial graph filtering implementation structure is presented to realize the proposed computation method. Finally, the station importance of Taipei metro system is identified by the conventional MI method and the proposed method. The top-K important stations are demonstrated to show both methods obtain the same results, so the proposed approximation method performs well.

In this paper, a deep metric learning method with combined loss of the triplet network and autoencoder is presented. Autoencoder is regarded as the regulation network to enable the embedding vector to have some latent features of the input image, and improve its performance. Compared with the pure triplet network, although it increases some complexity during training due to the addition of the decoder, but during testing, their complexities are exactly the same, because the decoder can be completely removed after training. The experiments of the proposed method, triplet network, and one-hot encoded network are performed on various character datasets to show that the proposed method not only achieve better classification performance, but also inherit the benefits of deep metric learning.

In this paper, graph signal denoising problem is investigated. First, conventional graph signal denoising method using graph Laplacian matrix (GLM) is described to show that a big matrix inversion is needed in this method. To reduce computational load, a modified Dijkstra's algorithm is presented to find the K-nearest neighbors (K-NN) of a given vertex in the graph and a local graph Laplacian matrix (LGLM) of the sub-graph around this vertex is constructed by using the K-NN information and graph adjacency matrix. Then, based on the local smoothness property of graph signal and LGLM, the denoised signal at the given vertex can be computed by a Cramer's rule method. Finally, real temperature data is used to show the effectiveness of the proposed denoising method and performance comparison with conventional method is made.

Image compression is a method to remove spatial redundancy between adjacent pixels and reconstruct a high-quality image. In the past few years, deep learning has gained huge attention from the research community and produced promising image reconstruction results. Therefore, recent methods focused on developing deeper and more complex networks, which significantly increased network complexity. In this paper, two effective novel blocks are developed: analysis and synthesis block that employs the convolution layer and Generalized Divisive Normalization (GDN) in the variable-rate encoder and decoder side. Our network utilizes a pixel RNN approach for quantization. Furthermore, to improve the whole network, we encode a residual image using LSTM cells to reduce unnecessary information. Experimental results demonstrated that the pro-posed variable-rate framework with novel blocks outperforms existing methods and standard image codecs, such as George’s [11] and JPEG in terms of image similarity. The project page along with code and models are available at https://github.com/khawar512/cvpr_image_compress

Under-display camera (UDC) technology is essential for full-screen display in smartphones and is achieved by removing the concept of drilling holes on display. However, this causes inevitable image degradation in the form of spatially variant blur and noise because of the opaque display in front of the camera. To address spatially variant blur and noise in UDC images, we propose a novel controllable image restoration algorithm utilizing pixel-wise UDC-specific kernel representation and a noise estimator. The kernel representation is derived from an elaborate optical model that reflects the effect of both normal and oblique light incidence. Also, noise-adaptive learning is introduced to control noise levels, which can be utilized to provide optimal results depending on the user preferences. The experiments showed that the proposed method achieved superior quantitative performance as well as higher perceptual quality on both a real-world dataset and a monitor-based aligned dataset compared to conventional image restoration algorithms.

2.5D IC packages are typically produced through Chip on Wafer (CoW) or Chip on Substrate (CoS) processes. Among these, 2.5D processing involves bonding the interposer chip and ASIC chip perpendicularly in sequence to the substrate for 2.5D package production. However, due to the large size of the interposer chip and its commensurately large high temperature warpage, attaching the interposer chip to the substrate carries a high risk of interconnection defects (non-wet/short) at the chip-substrate joints arising from the warpage difference. Moreover, the warpage difference between the attached interposer chip and ASIC also leads to a high risk of interconnection defects during bonding. For this reason, accurate prediction of the interposer chip high temperature warpage is of utmost importance in 2.5D packages In package warpage simulation, the reference (stress free) temperature is generally taken from high-T processes such as molding or F/C mounting, as they are most critical in inducing package warpage. However, unlike an IC package, the interposer chip undergoes no dominant process in particular which is critical to its chip warpage. Instead, all device fabrication processes prior to package formation impact the final interposer chip warpage. The initial room temperature warpage is determined by the residual stress from each processing step. The high temperature warpage is determined by the difference in metal density between the regions above and beneath the silicon as the temperature is increased. In the present study, we examine a $28.3\times 18.9\text{mm}$ 4-layer BEOL, 1-layer B-RDL interposer chip in order to predict the final interposer chip warpage via the finite element method (FEM) and thereby more effectively control it. To that end, we obtain the residual stress resulting from the front side (BEOL) and back side (B-RDL) processes through warpage measurement at each step. Moreover, we propose a method to maintain the interposer chip warpage close to zero, by adjusting the metal density and CVD film properties. In addition, we aim to address and improve possible Joint defects (non-wet/short) during ASIC chip bonding by predicting the warpage of the interposer chip attached to the substrate.

Advanced package technology has been developed rapidly to meet a demand of the high end application such as AI and datacenter. 2.5D silicon interposer technology has been focused as the solution, for high end applications because of its heterogeneous device integration compatibility: high bandwidth memories (HBMs), logic devices or functional chiplets. In this study, a 2.5D structure package called Molded Interposer on Substrate (MIoS) with an extremely large silicon interposer (>2800mm2) on the $85\times 85\text{mm}^{2}$ body size assembled with 2-ASICs and 8-HBMs was demonstrated successfully for higher chip integration capability. Also, the key challenges of extremely large size 2.5D MIoS package such as warpage of the molded interposer (MIP) module and high level of reliability subjected to thermo-mechanical stress were investigated. MIP warpage was simulated by finite element method (FEM) and controlled the warpage difference between MIP and substrate below 50um at solder melting temperature. As a result, the number of 60K bumps obtained the good joint quality during reflow bonding process. The package reliability was evaluated under thermal cycle test (−55∼125°C) for optimizing the stress induced by the mismatch of thermal expansion (CTE) of components: substrate, underfill, ring frame materials and epoxy mold compound (EMC). The primary failure modes were underfill crack and EMC crack at the corner of devices at the early stage but, through a study on components material properties, package level reliability was improved.

In this paper, three missing temperature data recovery methods using smoothness, bandlimitedness and sparseness priors are presented. First, the temperature data collected from the sensor network is represented by graph signal such that graph Laplacian matrix (GLM) and graph Fourier transform (GFT) can be employed to develop the missing data recovery methods. Second, the smoothness measure of graph signal is defined by GLM and the recovery problem based on smoothness prior is formulated as an optimization problem whose solution can be obtained by solving the matrix inversion. Third, a recovery method based on bandlimitedness prior in GFT domain is studied and an iterative method is used to get the recovery data. Fourth, the sparseness prior in GFT domain is applied to estimate the missing temperature data by the iterative thresholding method. Finally, real temperature data collected in Taiwan is used to evaluate the performance of three recovery methods based on different priors.

In this paper, the graph Fourier transform (GFT) centrality is used as node importance measure to identify central stations of the Taipei metro system. First, the computation algorithm of the GFT centrality is presented. Then, GFT centrality is employed to analyze Taipei metro system and determine which stations are more central. This information is useful to analyze the vulnerability of the metro system that considers the recovering ability after system suffers a disaster. Finally, the comparisons among degree centrality, closeness centrality, betweenness centrality, and eigenvector centrality are made to show their differences and characteristics.

In this paper, a temperature data denoising method using Laplacian matrix and Neumann series is presented. First, the denoising problem is formulated as an optimization problem whose solution needs to solve the matrix inversion. Then, to avoid solving matrix inversion, the Neumann series is used to approximate the matrix inversion by a truncated series expansion. Next, two efficient distributed implementation methods are presented to realize the proposed denoising algorithm. Finally, the real temperature data in Taiwan is used to demonstrate the effectiveness of the proposed method.

As faster data processing and communication gets more demanded for Data Center/Cloud, HPC (High Performance Computing), AI (Artificial Intelligence) accelerator and Network markets, HBM (High Bandwidth Memory) becomes main memory type to meet the required bandwidth performance. HBM integration with logic dies in a system level has been developed on 2.5D SiP (System in Package) Platform with Si Interposer having TSV (Through Silicon Vias) of which fabrication cost is rather high. Therefore, as the low cost solution, alternative 2.5D SiP Platform approaches such as Organic Interposer using Redistribution Layer (RDL) and Glass Interposer have recently been reported. In this paper, RDL Interposer package with 4 HBM and 1 logic is demonstrated as 2.5D package platform based on RDL-First Fan-out Wafer Level Package (FOWLP). The effect of RDL design factors on electrical performances is investigated using the eye diagram method and fine pitch multi-layer RDL structure (2um L/S RDL, 4 Layers) is designed accordingly. Fine pitch RDL process is established followed by the wafer level and unit level assembly processes and RDL Interposer package is confirmed to meet all the reliability requirements.

This paper presents a low-power stacked CMOS image sensor (CIS) in 65/14nm process. With 14nm process, we could achieve 29% less power consumption than the conventional CIS in 65/28nm process. The measured random telegraph noise (RTN) result shows the 65/14nm stacked CIS to guarantee the commercial sensor image quality even though a fin field-effect transistor (FinFET) process is three-dimension channel structure. The pixel array of the implemented chip consists of 12-mega pixels (Mp) with a dual-photodiode (2PD) in 1.4μm pixel pitch, and the sensor output provides 120 frames per second while it consumes 612mW for 12Mp image and 3Mp auto-focus data via a mobile industry processor interface (MIPI) physical layer (DPHY) up to 6.5Gbps/lane. The measured random noise is 2.2e-at 16× analog gain, and figure-of-merit (FoM) of analog-to-digital converters (ADCs) achieves 0.46e-nJ.

In this paper, the designs of graph filters using leastsquares (LS) method with parameter norm penalty are studied. First, the design problem of the graph filter which can be used to process irregular sensor network data is described. Then, the conventional LS method is employed to design graph filter and its numerical stability problem in high-order filter case is presented. Next, to solve numerical stability problem, the LS method with parameter norm penalty is applied to design the graph filter. Finally, the design examples are demonstrated to show the effectiveness of the proposed design method.

In this paper, the graph construction of signal denoising method using Laplacian matrix is investigated. First, the four-step procedure to obtain a graph signal representation is described. Then, the details of graph signal de-noising method using Laplacain matrix are presented. Next, four kinds of graph constructions with different connectivities are studied including Hamiltonian cycle, spanning tree, Delaunay triangle and complete graph. Finally, the real temperature data is used to evaluate the performance of the graph signal de-noising method and to show which graph construction is a better choice according to the value of signal to noise ratio (SNR).

In this paper, three types of convolution operations in convolutional neural networks (CNNs) are studied including regular convolution, separable convolution and group convolution. For regular convolution case, the modified VGG-19 is used to construct the deep networks. For separable convolution case, the MobileNet is applied to build deep model. For group convolution, the VGG-like plain network is used to construct the model. The experimental results of image classification on the CIFAR-10 and Flower102 datasets are used to evaluate the performance of CNNs and to demonstrate which convolution operation is a better choice according to accuracy and complexity.

In this paper, traditional one-hot encoded network is compared with three deep metric learning methods by using contrastive loss, triplet loss, and quadruplet loss on character classification task. Experimental results show that deep metric learning methods provide similar performance to one-hot encoded network, but have more advantages and less label requirements. On the other hand, by comparing these well-known deep metric learning methods under the same number of parameters, their differences are also summarized.

Peak shaving is an important technology in power grid. Many studies on peak shaving are being conducted, using ESS, shiftable load, and V2G. However, they do not consider of the ESS SoC and supply power reduction because they are only for peak shaving. ESS SoC is related to battery life which is directly linked to costs. Reduction supplying power causes problems that uncertainty and inconvenience to customers. Therefore, it is necessary to consider the problems of peak shaving. This paper presents ESS SoC optimization by controlling that charging and discharging of EV. It can reduce supplying power and peak shaving.

In this paper, a supervised learning method for the design of linear phase FIR digital filter using Keras is presented. First, the design problem of the linear phase finite impulse response (FIR) digital filter is transformed to a supervised learning problem. Then, the optimizers in Keras framework are used to determine the filter coefficients by minimizing the mean squared error (MSE) loss function. The widely-used optimizers include adaptive moments (Adam) algorithm and stochastic gradient descent (SGD) with momentum algorithm. Finally, the numerical design examples of low-pass and high-pass FIR digital filters are demonstrated to show the usefulness of the supervised learning method with Keras framework.

In this paper, the design of graph filter using spectral transformation and window method is presented. First, the specification of graph filter is converted to the one of digital filter using spectral transformation. Then, the window method is employed to design the digital filter with converted specification. Next, the binomial series expansion is used to obtain the coefficients of the graph filter from those of digital filter. Finally, the numerical design examples are demonstrated to show the effectiveness of the proposed graph filter design method.

This paper is about the development of mobile app for integrated control and real time monitoring of smart fish farm based on IoT. It controls an important role of water processing in the recirculating smart fish farm. The proposed smart fish farm consists of Water Tank1, Water Tank2, Balancing Tank and RAS. In this aquaculture the water flow process is regulated by the PID controller along with water level ultrasonic sensors, water temperature sensors, DO and pH sensors. This designed aquaculture system operates on manual mode, automatic mode and remote mode by using mobile phone. To verify the performance of the designed system, the field tests were conducted in Future Creative Technology Research Institute of Busan University of Foreign Studies. It showed satisfactory performance of smart fish farm by experimental results.

Since the sub 6 GHz spectrum supports the cell capacity and coverage that can provide diverse services, most of the network operators chose the corresponding spectrum in the 5G initial deployment stage. The digital beamforming with cross polarized array antennas is a promising technique to provide high data rate and cell coverage enhancement in a sub 6 GHz 5G NR system. Unfortunately, the uplink coverage at sub 6 GHz is more challenging because the mobile station beamforming is not possible due to the limited hardware area for embedding massive antennas and transmit power of mobile station is limited. Particularly, the uplink channel quality deterioration causes downlink beamforming performance degradation in the time division duplex (TDD) mode because the fully digital beamforming performs the downlink beamforming weight decision through the uplink channel estimation. We proposed the adaptive digital beamforming based on Cross Polarized Array Antennas and analyzed the performance for uplink coverage of cell edge in the sub 6 GHz 5G NR system in this letter.

The web design should not only deliver the basic merchandise trading, but also embellish the web site in order to enhance the purchase intention through the concept of visual aesthetics, color, pictures and films. Based on the theory Stimulus-Organism-Response(S-O-R), the model of aesthetic experiences, and two constructs of web visual aesthetics “aesthetic formality and aesthetic appeal”, this study establishes a reaction mode of web visual aesthetics perception. This study investigates the effect of user’s "web visual aesthetics perception" through "positive emotions", "aesthetic experience value" and "satisfaction" on "purchase intention". The software AMOS 20 is used to conduct a structural equation modeling analysis of the 288 respondents collected from the website Forestmosa. The result of data analysis validates a theoretical model that "web visual aesthetics perception" positively affects "positive emotions", "satisfaction" and "aesthetic experience value". "Positive emotions" positively affects "aesthetic experience value" and "satisfaction". "Aesthetic experience value" positively affects "satisfaction", and "satisfaction" positively affects "purchase intention". Also, the results show that regardless of the frequency of online shopping, improving the visual aesthetic perception of web pages can positively affect the aesthetic experience value. Especially, the visual aesthetic design of the shopping website's webpage is obviously important for online shopping low-frequency users.

Digital notch filter is a powerful tool to remove a sinusoidal or narrowband interference superimposed on a broadband signal. In this paper, the Chebyshev polynomial approximation method is employed to design linear phase FIR digital notch filter. Three main features of the proposed design are listed below: Firstly, the filter coefficients have closed-form solution without performing any optimization. Secondly, the notch filter has exact zero-gain response at the prescribed notch frequency. Thirdly, the 3-dB rejection bandwidth of notch filter can be adjusted by changing the filter order. Moreover, the power-line interference removal in electrocardiogram (ECG) signal is used to show the usefulness of the proposed notch filter.

In this paper, a fusion method for low-illumination color image enhancement on CIELUV space is presented. First, the input color image on RGB space is converted into the one on CIELUV space. Then, three color image enhancement methods are employed to enhance the low-illumination image individually including chroma-based method, homomorphic filtering method, and power-law transformation (PLT) method. Next, a Gaussian weighting fusion method is applied to combine three individual enhanced images to obtain the final enhanced image. Finally, low-illumination images are used to evaluate the enhanced quality of the proposed method and make comparison with conventional method.

In this paper, the design of digital differentiator (DD) using supervised learning method on Keras framework is presented. First, the design problem of DD is converted into a supervised learning problem such that the rich learning tools of the Keras can be applied to solve the problem directly. Then, the coefficients of transfer function of DD are determined by minimizing the loss functions such as mean squared errors or mean absolute errors. Next, the optimal solution of DD design problem is obtained by the various Keras optimizers such as stochastic gradient descent (SGD) or adaptive moments (Adam) algorithms. Finally, several numerical examples are shown to demonstrate the effectiveness of the proposed design method.

In this paper, a graph Fourier transform (GFT) based method for missing temperature data detection problem is presented. First, the frequency contents of GFT of temperature data are analyzed. If there is missing temperature data occurring in the sensor station, the high-frequency content increases in the spectral domain. Then, the energy ratio between high-frequency band and low-frequency band is used to develop a detector for detecting whether temperature data misses or not. The threshold value of detector is determined by computing the maximum energy ratio value of the regular daily data without missing temperature data. Finally, the real-world data measured from the weather sensor stations are used to evaluate the performance of the proposed detection method.

This paper presents a path tracking algorithm for autonomous driving that learns an action command from high-dimensional input state vector, e.g., grid-maps. The learning framework is built upon a variational auto-encoder (VAE) and takes advantage of the efficient path tracking results that already exist or can be obatined from a human expert. The VAE is known to give smooth latent represetations of the input data and we make one of the latent attribute follow the expert's command. We implement an autonomous driving system on the open source robot simulator (Webots) and collect the demonstration data for training the VAE. Numerical results show that the proposed tracking method can drive a vehicle robustly without explicitly detecting the road features.