Data collection and annotation is often a tedious and time consuming task to get satisfying image classification results. The transfer learning technique enables to overcome this challenge by reducing the number of images and the training time needed to add a new class. It is applied here to fire detection in the wild, detecting if there is a fire or no fire but can apply to many other situations.
Approach
– The tutorial presents how to use a technique called “Transfer learning” to quickly train a deep learning model to classify images.
– The tutorial is based on the computer vision function pack FP-AI-VISION1
Sensor
Vision: Camera module bundle (reference: B-CAMS-OMV)
Data
Dataset: Dataset for forest fire detection (License CC BY 4.0)
Data format:
2 classes: fire and no fire
RGB image 250x250x3
Results
Model: MobileNetV2 alpha 0.35
Input size: 128x128x3
Memory footprint:
403 KB Flash for weights
225 KB RAM for activations
Accuracy:
Float model: 98%
Quantized model: 98%
Performance on STM32H747 (High-perf) @ 400 MHz
Inference time: 112 ms
Frame rate: 8.9 fps
