Carrey Cho
Computer Vision Engineer
Profile
Computer Vision Engineer with 4+ years of experience in C++ vision library development, algorithm optimization, and industrial inspection systems. Specialized in object detection matching, high-density metrology. Patents and production deployment experience in semiconductor lines.
Experiences
Computer Vision Engineer
FourthLogic Inc.
FourthLogic Inc.
AI software company developing advanced computer vision and image processing platforms that combine classical algorithms and deep learning for industrial inspection systems. Current classical core algorithm performance is world-best performance, Achieved performance exceeding HALCON, which has long been th leading platform in industrial image processing. Sucessfully deployed in Hyundai Motor and Samsung Electronics production lines, improving inspection speed by 30% and reducing false positives rate to near 0%.
AI software company developing advanced computer vision and image processing platforms that combine classical algorithms and deep learning for industrial inspection systems. Current classical core algorithm performance is world-best performance, Achieved performance exceeding HALCON, which has long been th leading platform in industrial image processing. Sucessfully deployed in Hyundai Motor and Samsung Electronics production lines, improving inspection speed by 30% and reducing false positives rate to near 0%.
- As a early employee, contributed to classical core algorithm development with high performance. especially pattern matching and metrology
- Developed end-to-end and highly optimized C++ core algorithm for inspection pipelines.
- Designed memory-efficient and highly optimized architecture utilizing SIMD (AVX, SSE), multi-threading(furture, OpenMP) and CUDA.
- Refactored serialized structures into efficient parallel architectures.
- Design clean and modular architectures that promote scalability, long-term maintainability, and collaborative development.
- Developed high-performance general-purpose algorithms that extend beyond single-purpose solutions and are easy to apply across various image processing tasks.
- Managed development timelines and workflow management with DevOps practices.
- Architected independently operable and minimal algorithmic units to streamline debugging and significantly reduce issue diagnosis time
Skill
- Strong proficiency in C++ with stable structure and optimization
- Specialized in rule-based object detection, including NCC-based, shape-based and geometric matching
- Skilled in performance optimization: SIMD (AVX/SSE), multi-threading, CUDA, and memory efficiency
- Analyzing minimal hyperparameters for easy to use and general-purpose vision algorithm
- Solid foundation in linear algebra, statistics and data analysis
- Experienced in R, Python, PyTorch and TensorFlow
Patents
- US 12243286 B2 – Pattern Image Detection Method
- US 12094183 B2 – Geometric Pattern Matching Method and Device
- KR 10-25079270000 – Method For Generating Quadrangle Gauge
*Full list of 9 patents related to vision algorithms available upon request.
Education
UNIVERSITY OF SEOUL
Feb 2013 – Aug 2019
Bachelor of Statistics (Major), Bachelor of Economy (Double Major)
- Served as President of the Statistics Student Council, leading student initiatives and coordinating events
- Includes 21 months of mandatory military service (Dec 2013 – Sep 2015)
KOREA UNIVERSITY
Sep 2019 – Feb 2020
Master of Artificial Intelligence (Incomplete)
- Implemented NLP models from research papers using PyTorch and TensorFlow, from Word2Vec to Transformer
VISA
- Permanent Full working rights in Australia (No sponsorship required)
Projects
Geometric Pattern Matching
Designed a general-purpose vision pipeline across diverse object conditions with high performance and robustness.
Problem
robustness template match of occculsion handling, light reflection like gradation
Constraints
General-Purpose, Training requires only a single image, Minimizing hyperparameters, Easy to use, Intuitive API.
Approach
- Engineered a general-purpose vision algorithm designed to operate with minimal preprocessing across diverse object conditions
- Proposed a new vector-based geometric feature extraction framework for precise and interpretable structural modeling
- Improved detection accuracy by incorporating robustness to pixel-level perturbations and preserving geometric topological consistency
- Minimized hyperparameter dependency for users by automatically determining optimal internal parameters
- Designing a simplified interface, enabling intuitive and user-friendly operation
- Adpoted continuous refactoring and systematic testing practices to minimize maintenance overhead and accelerate feature deployment
Impact
*This algorithm is registered as a U.S. patent (US 12094183 B2).
Projects
Optimization
Optimized with SIMD (AVX/SSE) and multithreading.
Approach
- Refactored serialized structures into efficient parallel architectures
- Partitioned the system into sections to identify overhead and optimize performance efficiently within limited due time
- Enabled function-level parallelism by designing clearly separated and independent functions.
- Evaluated async and future functions for each section execution.
- Refactored code or decomposed large functions into smaller, more manageable units
- Test for caches-hit and memory access patterns included sharing memory and allocatation and so on
- Evaluated whether SIMD provided performance benefits under multi-threaded workloads
- Analyzed performance-to-resource efficiency to avoid unnecessary multithreading when SIMD alone was sufficient
- Compared SIMD instruction latency and throughput to guide optimization decisions
Impact
| Geometric Match | Find Image | ||
|---|---|---|---|
| Learn Image | Small (512x512) | Mid (2048x2048) | Large (5120x5120) |
| Small (32x32) | 0.8ms | 11.4ms | 22ms |
| Mid (128x128) | 1.5ms | 7.2ms | 12ms |
| Large (512x512) | - | 5.2ms | 8.9ms |
*Achieved 10–400% faster processing than HALCON across all cases