Edge AI Meets Microscopy

A good microscope allows us to view an entire world that is otherwise invisible to us. But to the untrained eye, this microscopic world can be difficult to understand. What looks like a blur of shapes and colors may actually be a highly organized ecosystem. Without knowing what to look for, a beginner might easily mistake something interesting for a dust particle. Of course, these skills can be learned, but that takes time.

As a shortcut, YouTuber That Project came up with the idea of building an AI-powered microscope. The basic idea behind it is that an AI algorithm could analyze images captured by the microscope to identify what is in focus. This wouldn’t replace human observers, but would give them extra information so that they can get more out of their time with the microscope.

To test the concept, the creator assembled a custom digital microscope around a DFRobot HuskyLens 2 AI camera and an NVIDIA Jetson Orin Nano. While the camera handles image capture and includes its own onboard machine vision capabilities, the Jetson serves as the computational powerhouse capable of running advanced AI models locally. Using this setup, the microscope can analyze specimens without relying on cloud services or an internet connection.

To make this work, the stock lens was removed from the HuskyLens 2 and replaced with a long microscope lens capable of providing roughly 30x magnification. A custom 3D-printed enclosure holds everything together and keeps the lens aligned with the image sensor. The design works, although the creator notes that achieving precise focus can be challenging because the current mount lacks a refined adjustment mechanism.

Connecting the hardware also required a bit of improvisation. Because of issues with the Jetson's serial interface, two separate USB connections were used. One cable carries control commands between the host computer and camera, while the second handles the live video stream. Once assembled, the microscope was ready for testing.

The first subject was pollen collected from a flower. The sample was placed on transparent tape and positioned over black paper to improve contrast. Since the microscope uses top-down illumination rather than the backlighting common in laboratory microscopes, the dark background helps highlight surface details that might otherwise disappear into the scene.

Images captured by the microscope were then passed to a locally running Gemma vision model. The AI was able to provide general observations about the specimen, but it quickly revealed the limitations of the setup when asked to identify the exact flower species. According to the model, the available magnification simply wasn't sufficient to reveal the microscopic structures needed for reliable identification.

The optics may need an upgrade, but some interesting features were still demonstrated. In one test, the HuskyLens 2's built-in object tracking system easily locked onto a tiny insect that wandered into view, following its movement across the frame in real time. The microscope also proved very useful for electronics work, allowing the creator to inspect tiny components and read markings on a Raspberry Pi Zero 2 W.

It may not replace a laboratory-grade microscope, but this AI-enhanced build shows how accessible hardware and edge AI can make microscopic exploration a little more approachable — especially with a few more improvements.