Exploring Feature Extraction Technique Parameters for Acoustic Gunshot Classification

The paper presents a systematic investigation of feature extraction techniques for acoustic gunshot classification, utilizing a dataset of 23,000 recordings from 85 firearms and 21 calibers. It benchmarks three feature extraction methods with 12 unique parameter sets using the ResNet-18 architecture, achieving a top-1 accuracy improvement of up to 20% with optimal techniques and an additional 4.7% with appropriate parameter tuning. This work is significant for practitioners as it highlights the importance of feature extraction choices in enhancing model performance in real-world gunshot detection applications.