mjs513
Senior Member+
As a further demo of the T4's capability I implemented a image classification using ARM's CMSIS-Neural Network with a ArduCAM for image capturing. The example uses the CIFAR-10 dataset that allows images to be classified into 10 groups ({"Plane", "Car", "Bird", "Cat", "Deer", "Dog", "Frog", "Horse", "Ship", "Truck"}) other datasets allow more flexibility. Ok enough of the overview.
In another post in the T4 Beta thread I posted on getting the ArduCAM software working on the T4. I posted that demo on the YouTube.
For this project I used a ST7789 2in 240x320 display along with the OV2640 ArduCAM (the T4 is mounted below the breadboard shield on a custom breakout board designed to accept Arduino type shields).
I created a Arduino style library for CMSIS-NN and incorporated it into the sketch I created to display the images. To make it work though I had to create a function to reduce the image size (320x240) bitmap down to a 32x32 image for the NN to work with CIFAR-10 dataset. The image I used for a test was of an airplane that I printed out (not very good print though as a test):
After I selected Bitmap and did a single capture using ArduCAMs interface a reduced size image along with the classification and confidence level is printed on the screen:
Note: I did test other images directly from the screen with out issue with the setup.
Thought you all you like to know what is possible with the T4.
In another post in the T4 Beta thread I posted on getting the ArduCAM software working on the T4. I posted that demo on the YouTube.
For this project I used a ST7789 2in 240x320 display along with the OV2640 ArduCAM (the T4 is mounted below the breadboard shield on a custom breakout board designed to accept Arduino type shields).
I created a Arduino style library for CMSIS-NN and incorporated it into the sketch I created to display the images. To make it work though I had to create a function to reduce the image size (320x240) bitmap down to a 32x32 image for the NN to work with CIFAR-10 dataset. The image I used for a test was of an airplane that I printed out (not very good print though as a test):
After I selected Bitmap and did a single capture using ArduCAMs interface a reduced size image along with the classification and confidence level is printed on the screen:
Note: I did test other images directly from the screen with out issue with the setup.
Thought you all you like to know what is possible with the T4.