Visual AI
MRO

Is it possible to inspect aircraft skin using mobile phone devices ? Yes!

AIVisualMRO is an advanced AI platform designed to help inspectors complete visual inspections more efficiently and consistently. It combines audio transcription, computer vision, and large language models to support every stage of the inspection workflow while ensuring that the human inspector always maintains full control over the final decision.

This one is the aircraft skin of Sea King HAR.3 Helicopter sitting int he RAF museum in London, I highly recomend visiting for 1/2 day.

Managed to take a video and train a AI model (computer vision model) to detect rivets and the status of them on the aircraft skin. This was done within 2 days, prelimiary training the AI model. You will see in a moment why it is important to do preliminary testing using AI models. It helps itterate fast and figure out if any problems will appear later in production.

On the main page showing a a video which is a small segment of the video inspection and video detection of rivets. But this does not tell the full story.

In this case there were 2 major problems, apart from the thousands of rivets that needed to be segmented / detected.

• - Small object detection: When the object you are trying to detect is much much smaller then the overall image then either the detection is "not stable" / flickering making it difficult to track in real-time or during recording. This results in errors in reporting or object tracking and identification. One potential solution might be to increase the size resolution image that the AI model is trained on, the other is to employ cameras with higher resolution. There is another solution to use an inferance slicer SAHI (link below as example), but that had limited success. It highly depends on case by case.


• - Flase positives: When detecting small objects in a bigger image, the AI model in this case was giving a lot of false positives, as you can see from the video below, especially at the begining and at the end of the video. However, Once the camera is close enough for the inspection, then the rivets are detected and inspected with high accuracy. Theese scenarios at teh begining and ending of the video were not accounted for in the AI training as the rivets were too small relative to the image size.

Below there is the full video. The artifacts at the begining and end of the inspection video were fixed in UI, starting the inspection video only when closer to the sircraft skin, reducing errors.

The above image is example of a validation dataset for training YOLO v11. I will make the full model available soon, for free. or contat us if you are interested.

In the meantime here are some resources that can be useful:

• How to Detect Small Objects: A Guide
• - A free python scripts to augment your own datasets
• RAF : Visit the RAF London museum !

Would you like to see for your self ? - Get in touch !

Every company performs visual inspection differently. AIVisualMRO begins with a non-intrusive observation phase to understand your current workflow, tools, constraints, and quality criteria. Requirements are gathered collaboratively to ensure the platform fits seamlessly into your existing operations.

Tailored to Your Inspection Process