Tsung‐Han Tsai

ORCID: 0000-0001-7524-0621
Publications
Citations
Views
---
Saved
---
About
Contact & Profiles
Research Areas
  • Advanced Data Compression Techniques
  • Video Coding and Compression Technologies
  • Digital Filter Design and Implementation
  • Advanced Vision and Imaging
  • Video Surveillance and Tracking Methods
  • Video Analysis and Summarization
  • Advanced Adaptive Filtering Techniques
  • Speech and Audio Processing
  • Image and Video Quality Assessment
  • Advanced Image and Video Retrieval Techniques
  • Human Pose and Action Recognition
  • Hand Gesture Recognition Systems
  • Music and Audio Processing
  • Image and Signal Denoising Methods
  • Algorithms and Data Compression
  • Analog and Mixed-Signal Circuit Design
  • Speech Recognition and Synthesis
  • Advanced Image Processing Techniques
  • Optical Coherence Tomography Applications
  • Image Enhancement Techniques
  • Industrial Vision Systems and Defect Detection
  • Advanced Neural Network Applications
  • demographic modeling and climate adaptation
  • Spatial and Panel Data Analysis
  • Image and Video Stabilization

National Central University
2016-2025

Ollscoil na Gaillimhe – University of Galway
2023-2024

National Cheng Kung University
2022-2023

Kaohsiung Chang Gung Memorial Hospital
2023

Chang Gung University
2023

META Health
2023

National Yang Ming Chiao Tung University
2022

National Taipei University of Technology
2019

ORCID
2019

Academia Sinica
2015-2016

For the semantic segmentation task, spatial information and receptive field are indispensable. to be practically applicable, it must have real-time inference speed. However, most of today's methods almost choose compromise resolution low-level detail information, which leads a significant decrease in accuracy. In this paper, we propose new architecture based on Bilateral Segmentation Network (BiSeNet) called BiSeNet V3. It introduces feature refinement module optimize map fusion combine...

10.1016/j.neucom.2023.02.025 article EN cc-by Neurocomputing 2023-02-16

In the semiconductor industry, testing section has always played an important role. The often requires engineers to judge defect, which wastes a lot of time and cost. accurate classification can provide useful information for through neural networks. this paper, we present method wafer map data augmentation defect classification. Data is based on CNN encoder-decoder depthwise separable convolutions. There are two datasets used, one open dataset WM-811K other built with Taiwan company. We...

10.1109/tsm.2020.3013004 article EN publisher-specific-oa IEEE Transactions on Semiconductor Manufacturing 2020-07-30

We study the problem of optimally adapting ongoing cloud gaming sessions to maximize gamer experience in dynamic environments. The considered is quite challenging because: 1) subjective and hard quantify; 2) existing open-source platform does not support reconfigurations video codecs; 3) resource allocation among concurrent gamers leaves a huge room optimize. rigorously address these three challenges by: conducting crowdsourced user over live Internet for an empirical model; enhancing frame...

10.1109/tcsvt.2015.2450173 article EN IEEE Transactions on Circuits and Systems for Video Technology 2015-07-21

In recent years, the fall detection system has become an important topic in homecare system. Compared with traditional algorithm, method used by neural network is more robust and higher accuracy. However consumes a large amount of energy due to huge number computations, needs memory store parameters as compared algorithms. this paper, we propose combination algorithm network. First, skeleton information extraction which transforms depth into extracts joints related activity. Also have...

10.1109/access.2019.2947518 article EN cc-by IEEE Access 2019-01-01

This paper presents a method for wireless ECG compression and zero lossless decompression using combination of three different techniques in order to increase storage space while reducing transmission time. The first technique used the proposed algorithm is an adaptive linear prediction; it achieves high sensitivity positive prediction. second content-adaptive Golomb-Rice coding, with window size encode residual prediction error. third use suitable packing format; this enables real-time...

10.1109/access.2018.2858857 article EN cc-by-nc-nd IEEE Access 2018-01-01

Hand segmentation aims to segment the hand profile however biggest challenge is over face or skin-related environments. To solve these problems, many previous papers rely on a very deep neural network collect new large-scale datasets real-life scenes increase diversity and complexity. perform it standard GPU, training inference time still long, always requires large amount of GPU memory. In this paper, we propose technique, Refined U-Net, based original U-Net [1]. The main objective with few...

10.1016/j.neucom.2022.04.079 article EN cc-by-nc-nd Neurocomputing 2022-04-20

With the outbreak of COVID-19, epidemic prevention has become a way to prevent spread epidemics. Many public places, such as hospitals, schools, and office require disinfection temperature measurement. To implement systems reduce risk infection, it is recent trend measure body through non-contact sensing with thermal imaging cameras. Compared fingerprints irises, face recognition accurate does not close contact, which significantly reduces infection. However, masks block most facial...

10.3390/s23062901 article EN cc-by Sensors 2023-03-07

At present, there are many researches on deep neural network (DNN) applied in life. In the task of object recognition, convolutional (CNN) has a good performance, but it relies GPU to solve large number complex operations. Thus hardware accelerator DNN is concerned by people. order implement model hardware, connection relationship and memory usage scheduling needed. This paper presnets design FPGA-based for DNN. The proposed architecture implemented Xilinx Zynq-7020 FPGA. It takes advantages...

10.1109/ddecs.2019.8724665 article EN 2019-04-01

In recent years, wireless vision sensor network (WVSN) is being used to retrieve video content from different image sensors which are connected devices wirelessly. This information do analysis can help automate tasks such as surveillance. For systems, power consumption during processing and communicating has been a challenge because of limited energy sources at the node. To deal with problem, in this paper WVSN proposed its algorithm hardware implementation for smart home application. The...

10.1109/access.2020.2982438 article EN cc-by IEEE Access 2020-01-01

In recent years, deep neural networks (DNNs) have brought revolutionary progress in various fields with the advent of technology. It is widely used image pre-processing, enhancement technology, face recognition, voice and other applications, gradually replacing traditional algorithms. shows that rise has led to reform artificial intelligence. Since network algorithms are computationally intensive, they require GPUs or accelerated hardware for real-time computation. However, high cost power...

10.1109/ojcas.2023.3245061 article EN cc-by IEEE Open Journal of Circuits and Systems 2023-01-01

In this paper, the VLSI-oriented fast, efficient, lossless image compression system (FELICS) algorithm, which consists of simplified adjusted binary code and Golomb-Rice with storage-less <i xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">k</i> parameter selection, is proposed to provide method for high-throughput applications. The reduces number arithmetic operation improves processing speed. According theoretical analysis, selection applies a fixed...

10.1109/tvlsi.2008.2007230 article EN IEEE Transactions on Very Large Scale Integration (VLSI) Systems 2009-03-30

10.1016/j.bspc.2020.101879 article EN Biomedical Signal Processing and Control 2020-03-25

Abstract A practical deep learning face recognition system can be divided into several tasks. These tasks time‐consuming if each task is executed with the original image as input data. And feature extractors used by different may duplicate its function. In this paper, a multi‐task training method based on pyramid and triplet loss to train single‐stage detection neural network proposed. As work, every task's data passed through same backbone avoid computation sharing weights computation. The...

10.1049/ipr2.12479 article EN cc-by IET Image Processing 2022-03-30

In recent years, intelligent video surveillance attempts to provide content analysis tools understand and predict the actions via sensor networks (VSN) for automated wide-area surveillance. this emerging network, visual object data is transmitted through different devices adapt needs of specific task. Therefore, they raise a new challenge delivery: how efficiently transmit various such as storage device, server, remote client server network. Object-based encoder can be used reduce...

10.1109/tmm.2011.2180705 article EN IEEE Transactions on Multimedia 2011-12-21

Sports video annotation, an active research area in the field of multimedia content understanding, is essential process applications, such as summarization, highlight extraction, event detection, and retrieval. This paper considers issue relation to annotation baseball videos. Conventional frameworks are based primarily on analysis, scoreboard recognition machine learning techniques, which require a substantial amount human input collect organize training data. The performance might become...

10.1109/tcsvt.2012.2189478 article EN IEEE Transactions on Circuits and Systems for Video Technology 2012-03-05

Motion estimation (ME) in the MPEG-4 AVC/JVT/H.264 video coding standard employs seven permitted block sizes to improve rate-distortion performance. This novel feature achieves significant gain over a macroblock using fixed size. However, ME is computationally intensive with complexity increasing linearly number of allowed sizes. paper presents an architecture for combined fast algorithm predict hexagon search (PHS) and edge information mode decision (EIMD). The EIMD utilizes best size...

10.1109/tcsvt.2011.2133230 article EN IEEE Transactions on Circuits and Systems for Video Technology 2011-03-29

Dynamic and adaptive binding between computing devices displays is increasingly more popular, screencast technologies enable such over wireless networks. In this paper, we design conduct the first detailed measurement study on performance of state-of-the-art technologies. Several commercial one open-source are considered in our analysis, which leads to several insights: (i) there no single winning technology, indicating rooms further enhance technologies, (ii) hardware video encoders...

10.1145/2713168.2713176 article EN 2015-03-09

An 8 b 700 MS/s 1 b/cycle asynchronous successive approximation register (SAR) analog to digital converter (ADC) which skips comparator metastability is presented. A delay-shift technique proposed shift the delay of generate a 1.5 redundancy range and accelerate comparison speed. This reduces settling requirement compensates for dynamic offset by redundancy. The prototype ADC in 40 nm CMOS technology achieves an SNDR 43.9 dB at Nyquist rate consumes 5 mW with 1.2 V supply. results FoM 56...

10.1109/tcsi.2016.2529278 article EN IEEE Transactions on Circuits and Systems I Regular Papers 2016-03-18

With the surveillance camera become more and popular, people need to observe a large number of monitor at one time. In this paper, we propose vision based indoor positioning for intelligent buildings solve problem. We capture frame make Gaussian mixture learning video background model. Then do subtraction get foreground object. After have object, track objects use direct linear transform integrate information bird's-eye map. finally all moving are traced displayed on view.

10.1109/igbsg.2016.7539419 article EN 2016-06-01

This brief presents a VLSI implementation of an efficient lossless compression scheme for electrocardiogram (ECG) data encoding to save storage space and reduce transmission time. As algorithm is able time, this opportunity has been seized by implementing memory-less design while working at high clock speed in VLSI. ECG comprises two parts: adaptive linear prediction technique content-adaptive Golomb Rice code. An low power presented. To improve the performance, proposed uses bit shifting...

10.1109/tcsii.2020.2978554 article EN IEEE Transactions on Circuits & Systems II Express Briefs 2020-03-06

Deep neural networks are widely used in computer vision applications due to their high performance. However, DNNs involve a large number of computations the training and inference phase. Among different layers DNN, softmax layer has one most complex as it involves exponent division operations. So, hardware-efficient implementation is required reduce on-chip resources. In this paper, we propose new fast activation function. The proposed hardware consumes fewer resources works at speed...

10.1109/aicas51828.2021.9458541 article EN 2021-06-06
Coming Soon ...