It implements a program that obtains previews from camera modules and performs hand-determined rock-paper-scissors.
It combines Arduino IDE sample sketches (camera, number_recognition) with NNC sample projects (hand-sign).
I clip & resize Preview, convert it to grayscale and load it into the model, and I can recognize the rock and paper, but I can't recognize the scissors.
The camera settings for Preview acquisition are as follows.
·5fps
·CAM_WHITE_BALANCE_INCANDESCENT
·Convert to CAM_IMAGE_PIX_FMT_YUV422->GRAY
·320*240->Central 224*224 to 28*28 Clip & Resize
When shooting, the camera is fixed 20cm above the white desk top, and the interior lighting (LED) + iPhone lights are used from near the camera.
Also, I would like to check the camera image, and I would like to know how to output Preview as well.
Please give us your opinion on the reason why scissors are not recognized in this setting.
spresense
Because the recognition results depend on the data set used for learning and the neural network designed, it is difficult to determine the cause.For example, if you have a "convolutional neural network" like LeNet and a sufficient amount of learning data from your environment, you will naturally improve your recognition rate.
LeNet has a sample project in the Neural Network Console.
Also
GitHub got the Spresense SDK, but I already answered with that does not contain dnnrt_hand_sign, but there is a sketch of the hand sign at the following URL:
As for Preview, there is an LCD driver in the compressed file above, so I think it can be diverted.
© 2024 OneMinuteCode. All rights reserved.