Rubbish detector

Getting started

• Clone repository

git clone https://github.com/luca-ant/rubbish_detector.git

• Install dependencies

sudo apt install python3-setuptools
sudo apt install python3-pip
sudo apt install python3-venv

• Create a virtual environment and install requirements modules

cd rubbish_detector
python3 -m venv venv
source venv/bin/activate

python3 -m pip install -r requirements.txt

Download dataset

Run the script download_dataset.py to download the dataset in the correct structure.

python download_dataset.py

Configuration

Change the file config.py and uncomment the model that you prefer.

Running

• Training: After choose the model in the config.py file, run the script train_rubbish_detector.py. You'll find training history and data in training_logs directory.

python train_rubbish_detector.py

• Evaluate: To evaluate whole model on test images and calculate accuracy run the script evaluate_rubbish_detector.py

python evaluate_rubbish_detector.py

• Predict: To use the classifier to predict a class for your image run the script predict_rubbish.py

python predict_rubbish.py PATH_TO_YOUR_IMAGE 

To use the webcam of your pc as input use the script camera.py and press SPACE BAR to take a photo and predict the the class.

python camera.py

Download already trained keras models

Download the zip archive containing all models and extract it in the main directory fo the repository. Use the following commands:

cd rubbish_detection
wget https://github.com/luca-ant/rubbish_detection/releases/download/models/models.zip
tar -xvf models_tflite.tgz

Training results

Keras model	Accuracy
InceptionV3	94.95%
MobileNetV2	92.55%
NASNetMobile	93.62%
ResNet50	94.68%

Convert to Tensorflow Lite and optimize

To convert the model into tflite version run the script convert_keras_to_tflite.py

python convert_keras_to_tflite.py

• Evaluate: To evaluate the non optimized tflite model run the script evaluate_rubbish_detector_lite.py

python evaluate_rubbish_detector_lite.py

• Predict: To use the non optimized tflite model to predict a class for your image run the script predict_rubbish.py

python predict_rubbish_lite.py PATH_TO_YOUR_IMAGE 

• Test: To test and compute accuracy of all optimized tflite models run the script test_tflite.py. You'll find the output in test_results_tflite directory.

python test_tflite.py

Download already converted and optimized tflite models

Download the tar archive containing all tflite models and extract it in the main directory fo the repository. Use the following commands:

cd rubbish_detection
wget https://github.com/luca-ant/rubbish_detection/releases/download/models_tflite/models_tflite.tgz
tar -xvf models_tflite.tgz

Optimization results

TFLite Model	Accuracy
InceptionV3	94.95%
InceptionV3 float16 quantization	94.95%
InceptionV3 weights quantization	94.41%
InceptionV3 integer quantization	94.41%
MobileNetV2	92.55%
MobileNetV2 float16 quantization	92.55%
MobileNetV2 weights quantization	34.04%
MobileNetV2 integer quantization	92.02%
NASNetMobile	93.62%
NASNetMobile float16 quantization	93.62%
NASNetMobile weights quantization	92.82%
NASNetMobile integer quantization	91.76%
ResNet50	94.68%
ResNet50 float16 quantization	94.68%
ResNet50 weights quantization	95.21%
ResNet50 integer quantization	94.68%

To be continued...

TensorFlow Lite models can run also on a smartphone such as Andorid or iOS. See here for details!