Full-stack development & data analysis
Front-end and back-end development, data analysis, and Linux server maintenance.
Angular · jQuery · JavaScript · PHP · R · CSS · HTML · Linux
干皓丞
Haocheng Kan
Computer vision for manufacturing & medicine
From medical image segmentation to industrial inspection lines — building AI systems that ship on real equipment.
Interview · Path
Two chapters, one through-line
Full-stack foundations first; then AI and semiconductor applications — with systems that connect software, hardware, and the factory floor.
AI & semiconductor applications
Generative AI, medical and industrial image segmentation, semiconductor defect inspection, automated visual inspection, and process optimization.
YOLO · U-Net · Mamba U-Net · RF-DETR · PyTorch · Python · PyQt · C#
Work & research
Medical research, industrial deployment
Academic work centers on medical image segmentation. Recent practice brings AI onto production lines for quality inspection, equipment integration, and maintenance.
- Image segmentation & object detection
- Synthetic data generation
- Cross-dataset / cross-line generalizability
- happykoala →
Engineering
Software + hardware + manufacturing
Beyond models: cameras, networks, GPUs, and root-cause troubleshooting on real systems.
Hardware experience
- Linux / Ubuntu server setup
- Computer & GPU installation
- Industrial camera & lighting setup
- Network devices, RJ45 / TCP/IP
- Equipment connection & verification
Troubleshooting process
- Check power and network first
- Verify camera and lighting
- Confirm image quality
- Test software and hardware settings
- Find root cause step by step
Manufacturing proof
- Industrial image inspection systems
- Production-line workflow simulation
- Quality inspection process
- ONNX deployment
- System integration and testing
Selected work
Tools that close the loop
From annotation and training GUIs to C# / Python line systems and modern detectors.
Open source
LabelMe contribution
Active work on the LabelMe annotation ecosystem — upstream collaboration and a maintained fork for industrial workflows.
Data pipeline
Annotation to trainable formats
Label with LabelMe, convert to YOLO / masks, and keep OK images in training so models learn the clean class as well as defects.
PyQt · Capybara
End-to-end training GUI
A thorough rewrite of LabelMe into a GUI for segmentation and detection: convert, augment, train, and export ONNX.
C# · ONNX
YoloIndustrySegmentation
Train YOLO to ONNX, then run inspection through a C# GUI built for production operators.
Python
Production line simulation
A Python module that simulates production-line behavior for safer iteration before hardware bring-up.
CV · YOLO & RF-DETR
Metal segmentation comparison
Side-by-side evaluation of classical detectors and RF-DETR on industrial metal surfaces.
Research
Segmentation models & publications
U-Net family ablations, qualitative results, and privacy-aware federated segmentation papers.
U-Net family
UNet · U2Net · TransUNet · VMUNet · ResUNet · UNet++ · UNet+++
Symmetric encoder–decoder design since Ronneberger et al. (2015), still the backbone of many medical and industrial segmentation stacks.
How results are scored
- mIoU
- Overlap between prediction and ground truth — higher is better.
- DSC
- Dice similarity for region agreement; common in medical imaging.
- Accuracy
- Pixel correctness; can look high on background-heavy data without reflecting true segmentation quality.
ICICS 2025 · CCF-C
OTTER
Optimized Training with Trustworthy Enhanced Replication via Diffusion and Federated VMUNet for Privacy-Aware Medical Segmentation.
WCCI 2026 · CCF-C
KoalaMamba
A Federated Method with Vision Mamba UNet for Privacy-Preserving Medical Segmentation.
Under review · Industrial
RABBIT
A Privacy-Aware Diffusion and VMUNet-Based Segmentation Framework for Multi-Factory Industrial Defects.
Next
Looking ahead
Linux systems programming, embedded devices, ONNX deployment, and foundation models such as DINOv3 — bridging research quality with factory reliability.
