Module.DSDN

DSDN

Dangerous Sound Detection Network module for gunshot detection using VGG16 and transformation to spectrograms

Folder structure:

• audio_set - samples of audio
• spec_set - samples of generated spectograms
• data - csv files from AudioSet site
• models - trained models and plots

Requirements

• numpy==1.19.5
• opencv-python==4.5.1.48
• tensorflow==2.4.0
• pytube==10.0.0
• skimage==10.0.0
• keras==2.4.3
• moviepy==1.0.3
• librosa==0.8.0

Required Data

To evaluate/train model, you will need to download the required datasets:

• Gunshot
• Fireworks
• Clapping
• Screaming

Or, you can easily extract all required data (transformed to spectrograms) using "extract_audio.py" module as below:

python extract_audio.py --labels_path data/labels.cvs --data_path data/data.csv --label [gunfire | screamming | clapping | ...]

Training

Training using ours settings:

python train.py --spectrograms_dir path//to//dir//

Example of training with different settings:

python train.py --s path//to//dir// -- resize 64 --epochs 100 --batch_size 32 --class_number 5

Prediction

To predict trained model you will have to load it with Keras:

model = load_model(MODEL_PATH)

and after data tranformation, evalute using predict method from Keras:

predict = model.predict(data)