In industries like manufacturing and infrastructure, even the tiniest visual change — a crack, rust, or missing bolt — can lead to major failures or losses. We noticed that manual inspections are slow, repetitive, and error-prone, especially under poor lighting or fatigue.
That’s when we asked ourselves:
“What if AI could act as a digital inspector — an extra pair of eyes that never tires?”
This thought inspired DeltaVision, an intelligent visual difference detection engine designed to see what humans miss and ensure reliability, safety, and quality
What it does
DeltaVision automatically compares two or more images of the same scene or object (taken over time) to detect and classify changes such as:
Surface wear or corrosion
Color fading or misalignment
Missing or displaced components
It highlights these differences using a heatmap overlay and generates a summary report that quantifies the change. In short — it helps humans make faster, more accurate inspection decisions.
How we built it
We built DeltaVision as a modular AI pipeline combining classical vision techniques with deep learning:
Preprocessing: Image alignment, noise removal, and lighting normalization using OpenCV.
Change Detection:
Used SSIM (Structural Similarity Index) for pixel-level comparison
SSIM(x,y)=(μx2+μy2+C1)(σx2+σy2+C2)(2μxμy+C1)(2σxy+C2) Applied ResNet feature extraction to detect subtle structural differences.
Classification: Lightweight ML model to label detected changes.
Visualization: Diff heatmaps rendered via Matplotlib + OpenCV overlays.
Frontend: Built a Streamlit dashboard for easy uploads and real-time viewing.
Backend: Flask API hosted locally for fast processing.
Challenges we ran into
Lighting inconsistencies between image pairs led to false positives — we had to normalize brightness and contrast.
Misalignment issues made even identical images appear different, requiring robust feature-based registration.
Dataset limitations: Very few labeled datasets existed for “change detection,” so we created synthetic ones.
Performance tuning: Balancing speed (SSIM) with deep accuracy (CNNs) was tricky within hackathon time limits.
Accomplishments that we're proud of
Built a fully working prototype capable of comparing images and generating visual diff maps.
Achieved over 90% accuracy in detecting real-world visual changes on test samples.
Designed a clean, intuitive dashboard for demos and inspections.
Created a system that is explainable and human-friendly, not just AI-driven.
Demonstrated how AI can support, not replace, human judgment in inspection tasks.
What we learned
That real-world computer vision involves much more than models — alignment, lighting, and data preprocessing matter just as much.
How to blend classical and deep learning methods to get better results.
The value of team collaboration — we split tasks efficiently, brainstormed solutions quickly, and kept iterating.
How important explainability and visualization are for user trust — people need to see what AI sees.
What's next for DeltaVision
Our team wants to take DeltaVision beyond a prototype:
Integrate real-time camera feeds to detect live changes.
Build a mobile app for on-site inspections.
Extend to 3D model comparisons for complex infrastructure monitoring.
Partner with manufacturing and audit organizations to test DeltaVision in real-world conditions.
Improve the engine with self-learning models that adapt to new environments automatically.
We believe DeltaVision can become a powerful tool for industries that depend on precision, safety, and quality — a true bridge between human intuition and AI accuracy.
Log in or sign up for Devpost to join the conversation.