LIAO TSEN YUNG Cover Letter
?
1 like?477 views
This cover letter summarizes Leo Liao's qualifications for computer vision and machine learning roles. He has over 10 years of experience developing computational vision applications, including leading a team that created printing press inspection products using CUDA-accelerated algorithms. He is skilled in areas like signal processing, computer graphics, computer vision, and machine learning techniques.
Convert to study guideBETA
Transform any presentation into a summarized study guide, highlighting the most important points and key insights.
![CURRENT PROJECT : Intelligent Baby Monitor Products
Chief Developer responsible for algorithms development and
ARM embedded system architecture ( Freescale
iMX.6 platform) based on Gstreamer pipeline framework
supporting live HLS/ WebRTC A/V streaming / combined with
pre-trained SVM/CNN model for purpose of crying detector, cute
face classi?er and on-line incremental SVM for cry
reason classi?er, illustrated as the following architecture
diagram.
Related Technology Lists:
LBP AdaBoost cascade Face Detector [ GLES 2.0 /GLSL ]
Face Type Classi?er Deep-learning [ CovNet VGG-D Model
]
Optimized Motion Detector [ GLSL Direct Texture / Mipmap
/ Stencil Test ]
Crying Detector and Crying Recognizer [ LaSVM MiniBatch
SMO and KKT incremental SVM solver ]](https://image.slidesharecdn.com/be015a86-1602-4168-b243-5bc6175894e0-150518043949-lva1-app6891/85/LIAO-TSEN-YUNG-Cover-Letter-3-320.jpg)
![Skin-Face for Sleeping-Detector [ Adaptive skin detection
with blob handling and moment-based
normalized comparison ]
H.264 Bit-stream Video Saliency and Tracking-Learning-
Detection for Baby In-Crib Analyzer [ video saliency map ]
HLS Live streaming [ based on gstreamer with nginx server
]
WebRTC P2P [ based on gstreamer + LibnIce ]
FINISHED PROJECTS
AOI INSPECTION [ AOI High-speed Printing Inspection/ AOI
Glueing Inspection / AOI Barcode Inspection ]
Applying CUDA programming tricks into several inspection
algorithms such as divide and conquer/transform and conquer
also as well reduce and conquer, try to solve most common
defeat inspection like foil, emboss, hologram, mismatch,etc.
AOI MEASUREMENT [ AOI Precision and Accuracy
Measurement ]
Applying randomized approximation and some statistical
estimation techniques into CUDA-version Fiducial Mark ( color
mark + cross-hair mark ) meanwhile develop sub-pixel accuracy
measurement module for length/curvature/size measurements .](https://image.slidesharecdn.com/be015a86-1602-4168-b243-5bc6175894e0-150518043949-lva1-app6891/85/LIAO-TSEN-YUNG-Cover-Letter-4-320.jpg)
![) ( GPGPU )
Next-Gen Dynamic Inspection ROI Generator ( GPGPU )
Nex-Gen ARM-based CUDA Toolkit ( Tegra K1 GPGPU )
[ CUDA FFT-BASED Image Registration ]
Build FFT-based image registration to reduce memory access
and computation loading form O(N^3) to O (Nl^2 logN ), speeding up
at lest ten fold compared to traditional Normalized Correlation
Coe?cient ( NCC ) method in our large size 512x512
template registration case and moreover develops FMM ( Fast
Marching Method ) in-painting algorithm to overcome blur-
sensitive and no user brush selection operation of vanilla version
FFT phase correlation method.
[ All-in-one Dynamic Inspection Region ( ROI ) Generator ]
Saving sample setting and allow di?erent mixing template during on-
the-?y inspection, build Segmentation algorithms so
do iterative growing region by seed ( label ) or doing like dual-form
contour ?tting by seed( label ).
All-in-one means there are several method related each other
but all in one principles
Segmentation = Classi?cation = Reconstruction ( Restoration in ill-
condition )
Categorized into two kind operation:
1. Active or Auto e.g. simply repeating original inspect ?lter twice (
two-pass CUDA . Our method implement stencil pattern
generator by QuickShift ( Mean Shift or Mode Seeking Clustering
) or stroke pattern generator by Minkowski sums, then Region-
Growing through CUDA Iterative Split/Merge coarse-to-?ne to do
dynamic parallelism kernel .
2. Semi-Auto with initial user interaction e.g. prior seed( label )
region given by initial hot-spot or lasso-kind tool, then repetitive
?nd more classi?cation region by means of CUDA GrabCut or](https://image.slidesharecdn.com/be015a86-1602-4168-b243-5bc6175894e0-150518043949-lva1-app6891/85/LIAO-TSEN-YUNG-Cover-Letter-7-320.jpg)
![Active Contour Algorithms.
[ Hologram-robust Fiducial Marks Registration ]
FAST with non-maximum suppression then followed by doing DTW
(Dynamic Time Warping) alignment matching in order to
overcome hologram e?ect ( varying illumination and
occlusion conditions ) and maintaining good corner-feature points
robust than edge-?lter, instead of shortcoming of correlation which
not stable with hologram illumination e?ect.
[ QCTool on Fiducial Color-mark (CUDA) : ColorToneMapping /
Canny /HoughLine/Circle / Rect Fitting / Location/ Color
Deviation / Boundary Defect decision ]
[ SR BarCode : Projection-based enhance 1-dim barcode
resolution to combat resolution ine?ciency ]
SKILLS, ACTIVITIES & INTERESTS:
Prize, Certi?cate, and Patent
1.a NISSAN Innovation Award : Panoramic Car-Vision
System
2008定圍臓?晩a廿?仟?Lp仝親室仟M?署p々誼麼僧尅
http://mobile.autonet.com.tw/cgi-bin/?le_view.cgi?
a8100057081003](https://image.slidesharecdn.com/be015a86-1602-4168-b243-5bc6175894e0-150518043949-lva1-app6891/85/LIAO-TSEN-YUNG-Cover-Letter-8-320.jpg)
LIAO TSEN YUNG Cover Letter
- 1. Cover Letter LIAO,TSEN-YUNG ( LEO ) 僧尅 Contact: E-mail : f92942099@gmail.com Mobile :+886 0976-787-839 SUMMARY OF QUALIFICATIONS 10+ years researches and developments to craft innovative e?cient and e?ective computational vision application.Specialized subjects include signal processing theory, multimedia software/hardware architecture, computer graphics, computer vision and modern machine learning. 3+ years lead a team of 5+ and successfully achieve numerous CUDA-accelerated algorithms enable printing press inspection products applied with high-speed 4K Line industry and 60 fps area camera. 5+ year industry expertise with broad range scope back to back from Embedded ARM systems to GPU Desktop and Cloud Computing. Well execute minimal instruction decision making of optimal design both balancing simplicity and complexity within domains of technology and management. High-par and goal-seeking self-driven developer. AREAS OF EXPERTISE: R&D manager and lead architecture experience at Masterworks AOI Algorithm and Embedded Team ( 2012.2 ~Present ) Automatic Optical Inspection ( AOI ) industry experience at Masterworks corp. ( 2012.2 ~Present ) IP-cam surveillance industry at Brainchild corp. ( 2010.2 ~2011.12 ) VALUE PROPOSITION :
- 2. Well familiar with problem-solving practical skills Top-Down v.s. Bottom-up approach Dynamic v.s. Static approach Sparse v.s .Dense storage or computation unit Soft v.s. Hard Decision Rule Multi-Scale v.s. Cross-Scope Numerous real-?eld experience on machine learning techniques SVM, Random Forest, Deep-Learning, Semi-supervised Learning and Re-Enforcement Learning OCR project Face Type Classi?er project Sound/ Music Finger-print project Visual Saliency context project PROFESSIONAL EXPERIENCE EDUCATION: Short-term Pilot Training Program ( US SF Bay Area ) Jan 2009 ~ Dec 2009 PhD Candidate Communication and Signal Processing Department TONIC Lab National Taiwan University ( R.O.C ) Feb. 2008 ~ Sep. 2009 MS. National Taiwan University ( R.O.C) Communication and Signal Processing Department Sep 2004 ~ Sep.2006 BS. EE National Taiwan University ( R.O.C) Sep.2000 ~ Sep.2004 WORKS & LEADERSHIP ACCOMPLISHMENTS:
- 3. CURRENT PROJECT : Intelligent Baby Monitor Products Chief Developer responsible for algorithms development and ARM embedded system architecture ( Freescale iMX.6 platform) based on Gstreamer pipeline framework supporting live HLS/ WebRTC A/V streaming / combined with pre-trained SVM/CNN model for purpose of crying detector, cute face classi?er and on-line incremental SVM for cry reason classi?er, illustrated as the following architecture diagram. Related Technology Lists: LBP AdaBoost cascade Face Detector [ GLES 2.0 /GLSL ] Face Type Classi?er Deep-learning [ CovNet VGG-D Model ] Optimized Motion Detector [ GLSL Direct Texture / Mipmap / Stencil Test ] Crying Detector and Crying Recognizer [ LaSVM MiniBatch SMO and KKT incremental SVM solver ]
- 4. Skin-Face for Sleeping-Detector [ Adaptive skin detection with blob handling and moment-based normalized comparison ] H.264 Bit-stream Video Saliency and Tracking-Learning- Detection for Baby In-Crib Analyzer [ video saliency map ] HLS Live streaming [ based on gstreamer with nginx server ] WebRTC P2P [ based on gstreamer + LibnIce ] FINISHED PROJECTS AOI INSPECTION [ AOI High-speed Printing Inspection/ AOI Glueing Inspection / AOI Barcode Inspection ] Applying CUDA programming tricks into several inspection algorithms such as divide and conquer/transform and conquer also as well reduce and conquer, try to solve most common defeat inspection like foil, emboss, hologram, mismatch,etc. AOI MEASUREMENT [ AOI Precision and Accuracy Measurement ] Applying randomized approximation and some statistical estimation techniques into CUDA-version Fiducial Mark ( color mark + cross-hair mark ) meanwhile develop sub-pixel accuracy measurement module for length/curvature/size measurements .
- 5. Accomplished Project Timeline 2012 Robust monotonic sequence algorithm on emboss defeat inspection ( CPU ) Orientation-range color ?ltering for inspection ( GPGPU ) Sub-pixel measurement via Facet Model/Gaussian N-jet model ( CPU ) 2013 CNY OCR module: Lightweight DSP SVM / LDA ( CPU )
- 6. QR/DataMatrix Decoder and Quali?cation ( CPU + GPGPU ) Build CUDA Test-suite Validation ( Precision and Regression Unit Test) for AOI integration on CUDA 5.5 ( GPGPU ) CUDA 5 pro?le-based optimization ( GPGPU ) Reduce memory tra?c and meet peak BW, balance between TLP and ILP in CUDA kernel fusion way. More CUDA Streams asynchronous operation / Texture object Design CUDA memory pool ( Object pool ) support for reduce time CudaMalloc time on-the-?y. FFT-based Image Registration ( GPGPU ) Bilateral NLM De-noising pre-processing ( GPGPU ) CUDA Run-length compression/decompression for saving global memory in 8K * 12K REGION MAP Inspection ( GPGPU ) Transparent Glue-over/less Visual Inspection (CPU ) Integration with open-source CUB project helps more device- level, block-level, warp-level abstraction and APIs reuse. ( GPGPU ) 2014 Dim defeat inspection on blank Full-Sheet ( GPGPU ) Texture analysis/synthesis Two-pass module for pattern full- sheet ( GPGPU ) Context analysis: Compressed domain Video Saliency Map ( GPGPU + CPU ) QC tool: Color ?ducial mark/ Camera + Line Scan Camera FOV Pose / Belt Line Vibration / Benchmark Test ( GPGPU +CPU ) Shape-context OCR module : TPS and K-D tree ( GPGPU ) Sub-pixel Measurement ( CPU ) Super-Resolution Bar-code Decoder ( CPU ) FAST-corner features plus with DTW ( Dynamic Time Warping) matching to combat common hologram e?ect ( CPU ) Next-Gen Textile Inspection: CUDA Dynamic parallelism Kernel ( GPGPU ) Next-Gen Stereoscopic Inspection : CUDA KinectFusion ( SLAM
- 7. ) ( GPGPU ) Next-Gen Dynamic Inspection ROI Generator ( GPGPU ) Nex-Gen ARM-based CUDA Toolkit ( Tegra K1 GPGPU ) [ CUDA FFT-BASED Image Registration ] Build FFT-based image registration to reduce memory access and computation loading form O(N^3) to O (Nl^2 logN ), speeding up at lest ten fold compared to traditional Normalized Correlation Coe?cient ( NCC ) method in our large size 512x512 template registration case and moreover develops FMM ( Fast Marching Method ) in-painting algorithm to overcome blur- sensitive and no user brush selection operation of vanilla version FFT phase correlation method. [ All-in-one Dynamic Inspection Region ( ROI ) Generator ] Saving sample setting and allow di?erent mixing template during on- the-?y inspection, build Segmentation algorithms so do iterative growing region by seed ( label ) or doing like dual-form contour ?tting by seed( label ). All-in-one means there are several method related each other but all in one principles Segmentation = Classi?cation = Reconstruction ( Restoration in ill- condition ) Categorized into two kind operation: 1. Active or Auto e.g. simply repeating original inspect ?lter twice ( two-pass CUDA . Our method implement stencil pattern generator by QuickShift ( Mean Shift or Mode Seeking Clustering ) or stroke pattern generator by Minkowski sums, then Region- Growing through CUDA Iterative Split/Merge coarse-to-?ne to do dynamic parallelism kernel . 2. Semi-Auto with initial user interaction e.g. prior seed( label ) region given by initial hot-spot or lasso-kind tool, then repetitive ?nd more classi?cation region by means of CUDA GrabCut or
- 8. Active Contour Algorithms. [ Hologram-robust Fiducial Marks Registration ] FAST with non-maximum suppression then followed by doing DTW (Dynamic Time Warping) alignment matching in order to overcome hologram e?ect ( varying illumination and occlusion conditions ) and maintaining good corner-feature points robust than edge-?lter, instead of shortcoming of correlation which not stable with hologram illumination e?ect. [ QCTool on Fiducial Color-mark (CUDA) : ColorToneMapping / Canny /HoughLine/Circle / Rect Fitting / Location/ Color Deviation / Boundary Defect decision ] [ SR BarCode : Projection-based enhance 1-dim barcode resolution to combat resolution ine?ciency ] SKILLS, ACTIVITIES & INTERESTS: Prize, Certi?cate, and Patent 1.a NISSAN Innovation Award : Panoramic Car-Vision System 2008定圍臓?晩a廿?仟?Lp仝親室仟M?署p々誼麼僧尅 http://mobile.autonet.com.tw/cgi-bin/?le_view.cgi? a8100057081003
- 9. 1.b Grant Patent : l祇庁M、宥?圭隈式凪X殻塀a瞳 Taiwan Patent 097146588 (??利購庁翠式凪宥儷?圭隈 CN 101754483 B) Gateway module, communication method, and computer program product l祇庁M、宥?圭隈式凪X殻塀a瞳 Application Number: 097146588 http://www.google.com/patents/CN101754483B?cl=zh 2.a Secutech 2011 IP HD/Megapixel Top Picks Cam http://www.secutech.com/11/download/news/55926.4375.pdf 2.b Co-partner Startup in App-works Venture Capital 3. CUDA Certi?cated NO.201309-01-0009 Programming Skills C/C++ Programming CUDA Programming Model ( evolved with CUDA generations from Fermi , Kepler and new Maxwell. ) Node.js Matlab Numeric Analysis Design Pattern Expertise Topics Modern Machine Learning Modern Computer Vision Algorithms Modern Embedded System Architecture Pipelines & Plug-ins Multimedia Systems Architecture
- 10. Modern Networking Architecture and Protocol Stack Innovative algorithm design and analysis