Image Recognition¶

Hold your objects in front of the Reef's camera for a few seconds each, train on the computer and deploy to the board. From then on the board's NPU classifies whatever is in the box at the center of the camera view and draws the answer on the camera stream. It all happens on IPRO AI Studio's Teach page, with no code.

Prerequisites¶

• An IPRO7AI Reef EVK flashed with the Edge AI firmware; see Flash the Edge AI Firmware.
• The OV5640 camera module: plug its cable into FPC1 with the board unpowered, and set SW7 to CAM. Always unplug USB-C before plugging or unplugging the camera cable.
• A USB-C data cable from the Reef's USB-C to the computer.
• IPRO AI Studio installed and allowed to use the camera; see Install IPRO AI Studio.

How it works¶

• The OV5640 captures 1280×720 and the ISP makes two outputs at once: a 640×360 preview, compressed to MJPEG and sent to the computer as a standard USB camera (UVC), and the box at the center, scaled to 128×128 for the NPU.
• The box at the center of the preview is exactly what the model sees. The Teach page cuts the same box out of the preview as training pictures ("crops").
• The model is a frozen MobileNet backbone plus a classification layer. Only the last layer is trained, so training takes seconds on the computer's CPU. Every layer runs on the NPU.
• On a deploy, Studio sends the model and the class names over the USB serial port. The board checks the size, CRC-32 and model format, writes the model to flash only if they are right, and restarts by itself.
• The board draws its result outside the box. The inside of the box always stays the camera's raw image, because the Teach page cuts its training pictures from there.

Connect the board¶

The Teach page¶

Create classes and record¶

Tips for recording:

• Bring small objects close to the lens so they fill most of the box.
• Turn the object slowly and move it closer and farther while recording, so the model sees different angles and distances. Two classes that look alike may be confused at some angles; record more at those angles.
• Record in the light and against the background you will use. Training adds a mirrored copy and two brightness changes of every crop by itself; you need not record mirrored views.
• The preview is not mirrored by default: you see what the camera sees. To move it like a mirror, tick Mirror; it does not affect training.

Train¶

One crop in every four of each class is held out of training and used to score the model afterwards: that is Held-out accuracy. Crops from one recording look alike, so the score says the classes can be told apart, not that a new scene will be right. With too few crops to hold any out, the page says Too few crops to hold any out for checking.

If the result says <A> taken for <B> (<n>x), n crops of class A were classified as B - usually because the object left the box while recording. Press Review on both classes, drop those crops and train again.

The NPU times in the menu are approximate. INTELPRO measured about 10.5 ms per run for MobileNetV2 0.35 on the board's NPU with Edge AI firmware 2.0.0.

Deploy¶

The board checks a model before writing it. If it refuses one, the model already on the board is untouched and you can deploy again straight away. If power or the cable is cut while writing, the image tool - which has a single slot - is left without a model and the camera view shows NO MODEL: DEPLOY INTERRUPTED; deploy again.

See the result on the board¶

After the deploy, put an object in the box. The board's NPU classifies it live and draws the result on the camera stream:

The Reef draws this into its own USB camera stream, so any program that can open a USB camera shows it, not only Studio.

You can also check the state on the console:

Verify¶

1. With each class's object in the box, its name shows left of the box, and its probability on the right is the highest and in green.
2. With the box empty, the result is BACKGROUND.
3. The class count at the bottom of the view matches the Teach page.
4. After unplugging USB-C or a restart, the model is still on the board: the view no longer shows NO MODEL, and Put it on the board shows Running <n> classes: ….

Add samples and retrain¶

The training score only says the collected crops can be told apart. Test in the real scene; where the board gets it wrong, record more at that angle or in that light, then train and deploy again:

1. Select the class that was misread, hold Hold to record at the angle and light where it went wrong, and press Review to drop bad crops if needed.
2. Press Train. After the classes or crops change, the training area says The crops changed since training. Train again.
3. Press Deploy.

Keep and reuse what you collected:

• Save…: choose a folder; Studio writes one sub-folder per class with the crops as PNG files.
• Open…: open a folder of the same layout (one sub-folder per class, .png or .jpg) to add samples or retrain without collecting again. 128×128 pictures are used as crops; other sizes are treated as camera frames and the center box is cut out. Folder names become class names.

• The same folder can be trained from the command line, without the window, into a model package and deployed:

  ipro-ai train image <folder> --out cups.ipkg --backbone mnv2_035
  ipro-ai deploy cups.ipkg --switch
  

  --backbone is mnv2_035 (MobileNetV2 0.35) or mnv1_025 (MobileNetV1 0.25).

Specifications¶

Tips¶

• Always have a BACKGROUND class. Without it, the board still picks one of the known classes when the box is empty.
• Image recognition has no "I don't know" answer. An object it was not taught goes to the most similar class, possibly with high confidence. For other objects that will appear in view, add a class (for example OTHER) and record them.
• Only what is inside the box counts; an object that is only outside the box is not recognized.

Troubleshooting¶

For anything else, see Troubleshooting.

Next steps¶

• Voiceprint Recognition: teach the same Reef who is speaking.
• Image Teach in the Browser: the same flow in Chrome or Edge, without Studio.