#covid19 #stm32 #ai
In this project, a Keras model for face mask detection is developed. We use a pruned model and then the project runs on a Nucleo Development board(F746ZG).
This is a sample output of the code that run on PC "Mask Detection.ipynb" and "eval_FM.ipynb":
source image : https://s.abcnews.com/images/Politics/trump-michigan-15-rtr-jc-200521_hpMain_16x9_992.jpg
after importing "my_model_f.h5" model to embedded platform we have some difference in the output.
P image : sample images that have mask
N image : sample images that have not mask
The output of system on PC:
network output for N11 result for N11.bmp is without_mask
network output for N12 result for N12.bmp is without_mask
network output for N13 result for N13.bmp is without_mask
network output for N14 result for N14.bmp is without_mask
network output for P11 result for P11.bmp is with_mask
network output for P12 result for P12.bmp is with_mask
network output for P13 result for P13.bmp is with_mask
network output for P14 result for P14.bmp is with_mask
| image | probability of no mask in image | probability of mask detection |
|---|---|---|
| P11 | 0.9862385 | 0.01376154 |
| N11 | 2.6498967e-06 | 9.9999738e-01 |
| P12 | 0.9984925 | 0.00150748 |
| N12 | 9.114358e-04 | 9.990885e-01 |
| P13 | 0.9965546 | 0.00344538 |
| N13 | 0.03975931 | 0.9602407 |
| P14 | 9.999670e-01 | 3.304353e-05 |
| N14 | 0.01601863 | 0.9839813 |
The output of system on Nucleo board:
| image | probability of no mask in image | probability of mask detection |
|---|---|---|
| P11 | 0.000309 | 0.999691 |
| N11 | 0.999966 | 0.000034 |
| P12 | 0.000158 | 0.999842 |
| N12 | 0.000334 | 0.999666 |
| P13 | 0.000334 | 0.999665 |
| N13 | 0.999966 | 0.000034 |
| P14 | 0.000158 | 0.999842 |
| N14 | 0.000308 | 0.999692 |
The main source code for reference and datasets
such a model can use in the street and with a red or green led signal to the pedestrian based on face mask status. for completion, it needs a face detection model that runs on board before mask detection. Now the face detection is not developed on board.
these models could be used with an attendance system for face mask detection.
model file : my_model_f.h5
type : keras (keras_dump) - tf.keras 2.4.0
c_name : network
compression : 4
quantize : None
model_name : my_model_f
input : input_0 [10000 items, 39.06 KiB, ai_float, FLOAT32, (100, 100, 1)]
inputs (total) : 39.06 KiB
output : dense_1_nl [2 items, 8 B, ai_float, FLOAT32, (1, 1, 2)]
outputs (total) : 8 B
params # : 107,872 items (421.38 KiB)
macc : 6,238,250
weights (ro) : 115,112 B (112.41 KiB) (-73.32%)
activations (rw) : 207,760 B (202.89 KiB)
ram (total) : 247,768 B (241.96 KiB) = 207,760 + 40,000 + 8
my_model_f p=107872(421.38 KBytes) macc=6238250 rom=112.41 KBytes (-73.32%) ram=202.89 KiB io_ram=39.07 KiB