VisionGuard AI

Inspiration

Industrial inspection often requires people to repeatedly examine equipment images and video footage to identify unusual visual patterns. Manual inspection can be time-consuming, while basic computer vision systems often stop after detecting an object or highlighting a suspicious region.

We wanted to explore a more intelligent approach: what if a vision system could analyze visual evidence, decide what to investigate next, and explain its next action?

This idea inspired VisionGuard AI, an autonomous visual inspection prototype built around OpenCV and agentic decision-making.

What it does

VisionGuard AI analyzes uploaded industrial images and videos using computer vision techniques. It identifies candidate regions that may require further inspection, visualizes the analysis, and uses an agent-style decision policy to recommend the next investigative step.

The system follows a simple loop:

See → Investigate → Decide → Act

It can recommend examining a candidate region more closely, reviewing additional video frames, or escalating an uncertain result for human review. Candidate regions are treated as inspection leads, not confirmed defects.

How we built it

We built VisionGuard AI as a Python-based Streamlit application with OpenCV at its core.

The application combines grayscale conversion, Gaussian blurring, Canny edge detection, morphological operations, contour analysis, and candidate-region ranking. For video inputs, it samples frames and analyzes them to identify areas that may warrant additional investigation.

A deterministic agent policy uses the visual analysis results to select a follow-up recommendation. The dashboard presents annotated images, inspection results, and a decision trace. It can also export a structured JSON report, with optional Amazon S3 integration for storing inspection artifacts.

Our design emphasizes explainability and a practical demonstration of how computer vision can guide subsequent actions rather than simply produce an image-processing output.

Challenges we faced

One of the main challenges was distinguishing potentially interesting visual regions from genuine defects. Edges, contours, shadows, and unusual textures can indicate areas worth investigating, but they do not independently prove that equipment is damaged.

Another challenge was designing a meaningful agentic workflow without overstating the system's capabilities. We addressed this by making the next-action policy explicit and preserving the distinction between detected visual candidates and verified defects.

Dependency compatibility, video processing, sensitivity tuning, false positives, and optional cloud integration also require careful testing. These challenges reinforced the importance of reproducible experiments, clear limitations, and human oversight.

What we learned

This project helped us understand how classical computer vision can serve as the perception layer of an agentic system. We learned to connect visual evidence with explicit decision rules, structure an inspection workflow, generate machine-readable reports, and plan evaluation around accuracy, false positives, and processing latency.

VisionGuard AI represents a foundation for more advanced inspection agents that could later incorporate trained defect classifiers, reference-image comparison, temporal analysis, and validated AWS deployment.

Future scope

Future improvements include evaluating the system on labeled industrial datasets, integrating a trained anomaly-detection model, benchmarking OpenCV 5 performance, and adding a feedback mechanism for human inspectors.

Our goal is to move toward visual inspection systems that not only identify where something may be wrong, but also explain what they should examine next.

Built With

• agentic-ai
• amazon-web-services
• artificial-intelligence
• boto3
• canny-edge-detection
• computer-vision
• contour-detection
• explainable
• image-processing
• industrial-inspection
• opencv-5
• python
• streamlit
• video-analytics