Hardware
- GPU-accelerated AI training systems
- Industrial imaging datasets
Computer Vision
Industrial Image and Video Dehazing Research for Autonomous Vision Systems
Conducted industrial AI research on image and video dehazing using transformer-based architectures, diffusion-assisted reconstruction, optical flow, and hybrid dual-domain neural networks during a Mitacs internship at MacDon Industries.
Summary
Engineering context
Conducted industrial AI research on image and video dehazing using transformer-based architectures, diffusion-assisted reconstruction, optical flow, and hybrid dual-domain neural networks during a Mitacs internship at MacDon Industries.
- Category
- Computer Vision
- Year
- Jan 2025 - Jan 2026
- Status
- Research Internship
- Context
- Mitacs @ MacDon Industries (Jan 2025 - Jan 2026)
My Role
Computer Vision and AI Research Engineer
Technical Stack
System Architecture
- Dual-domain transformer architecture processed spatial and frequency-domain image information
- Video dehazing pipeline integrated temporal consistency mechanisms
- Stable Diffusion and Video Stable Diffusion assisted patch reconstruction workflows
- Optical-flow-guided moving frame windows improved temporal reconstruction consistency
- AI perception systems targeted industrial autonomous-vision applications
Engineering Challenges
- Maintaining temporal consistency in video dehazing
- Reconstructing degraded image regions reliably
- Integrating diffusion-based reconstruction into restoration pipelines
- Handling moving particles and dynamic scene degradation
- Improving dehazing quality for autonomous-vision applications
Hardware / Firmware / Software
Software
- PyTorch
- Transformer architectures
- Stable Diffusion
- Video Stable Diffusion
- Optical flow pipelines
- AI research frameworks
Sensors
- Industrial imaging systems
Results / Outcomes
- Developed hybrid dual-domain transformer dehazing architectures
- Developed image and video dehazing pipelines
- Integrated diffusion-assisted patch reconstruction systems
- Implemented optical-flow-guided temporal reconstruction workflows
- Conducted industrial AI research under Mitacs internship collaboration
- Two research papers submitted for peer review
Engineering Notes
Dual-Domain Transformer Architecture
A hybrid dual-domain transformer architecture was developed to process:
- spatial-domain image information
- frequency-domain image information
The architecture targeted improved restoration of degraded industrial imagery and enhanced visual clarity for autonomous-vision systems.
Video Dehazing and Temporal Processing
Video dehazing pipelines were developed using:
- optical flow
- moving-frame temporal windows
- temporal consistency mechanisms
The system attempted to preserve temporal stability while restoring degraded video sequences.
Diffusion-Assisted Reconstruction
Stable Diffusion and Video Stable Diffusion were integrated into reconstruction workflows for:
- degraded-region reconstruction
- patch restoration
- image enhancement
- video refinement
The diffusion-assisted approach was explored for improving reconstruction quality in heavily degraded scenes.
Industrial AI Research
The research targeted industrial autonomous-vision applications where airborne dust and haze reduce perception quality.
Experiments were conducted using industrial datasets including RVID-based evaluation workflows.
Research Outcomes
The project resulted in:
- multiple experimental AI architectures
- industrial image-restoration pipelines
- video dehazing systems
- temporal reconstruction workflows
Two research papers related to the work were submitted for peer review.
Related projects
More engineering case studies in this domain
Autonomous Pipe Inspection FPV Drone System
Developed an autonomous confined-space FPV drone system for pipe inspection using PX4 firmware modification, optical-flow fusion, laser-based centering, and YOLO-based defect segmentation.
Autonomous Drone System for Copper Cable Recovery and Infrastructure Inspection
Developed an autonomous PX4-based drone platform for utility-infrastructure inspection and copper-cable recovery using Mask R-CNN segmentation, monocular metric depth estimation, and vision-guided autonomous positioning.
Autonomous Crawlspace Inspection Robot
Developed an autonomous crawlspace-inspection robot using ROS2, Nav2, SLAM Toolbox, frontier exploration, AI-based hazard detection, and remote iOS-based operation.