Deep Spectrogram Learning for Gunshot Classification: A Comparative Study of CNN Architectures and Time-Frequency...

Area of Science:

• Acoustics
• Machine Learning
• Signal Processing

Background:

• Gunshot sound classification is vital for public safety, forensic analysis, and surveillance.
• Deep learning models offer potential for improving the accuracy and efficiency of firearm sound identification.

Purpose of the Study:

• To evaluate deep learning model performance for classifying firearm sounds.
• To investigate the effectiveness of various time-frequency spectrogram representations converted into images for CNN analysis.

Main Methods:

• Analyzed twelve time-frequency spectrograms (Mel, Bark, MFCC, CQT, Cochleagram, STFT, FFT, Reassigned, Chroma, Spectral Contrast, Wavelet).
• Converted spectrograms into RGB images for application of computer vision techniques.
• Trained six Convolutional Neural Network (CNN) architectures (ResNet18, ResNet50, ResNet101, GoogLeNet, Inception-v3, InceptionResNetV2) on spectrogram images.

Main Results:

• CQT, Cochleagram, and Mel spectrograms achieved high classification accuracy (>94%) when used with deep CNNs like ResNet101 and InceptionResNetV2.
• Transforming spectrograms into images enabled effective use of image-based processing and deep learning models.
• The approach demonstrated robustness in capturing spectral-temporal patterns for accurate firearm sound classification.

Conclusions:

• Deep learning models, particularly CNNs, combined with image-based spectrogram analysis, provide a powerful framework for firearm sound classification.
• Specific spectrogram representations (CQT, Cochleagram, Mel) are highly effective when converted to image formats for deep learning.
• This methodology enhances accuracy and robustness in applications like public safety and forensic investigations.