Evaluation of Image Classification Networks on Impulse Sound Classification Task

Impulse sound classification has several civil and security related applications. Convolutional Neural Networks (CNN s) have been proven very effective in image classification and show promise for audio applications as well. The main aim of this paper is to see if we can use pre-defined image classification networks for impulse sound classification task and evaluate their performance on said task. We use ballistic sound dataset containing gunshots of different weapons for impulse sound classification task. For using audio segments as images, we calculate a spectrogram representation of the data and treat them as images for the training. A convolutional 2D task specific network, SimpleNet 2D and a convolutional 1D task specific network, SimpleNet 1D with similar architecture are designed and their performance is compared to study the effect of data representation on training. From state-of-the-art of image classification networks, VG-GNet, Inception v3, Inception-ResNet- v2, NASNet, Res NeXt, EfficientNet are chosen as candidate architectures for evaluation. These networks along with SimpleNet 2D are evaluated on four different test datasets and we compare the performance of the networks when trained from scratch and trained on ImageNet pre-trained weights. From these evaluations, the effect of pre-trained weights on training is observed.