Surgical Mask Detection with Convolutional Neural Networks and Data Augmentations on Spectrograms

Abstract

In many fields of research, labeled data-sets are hard to acquire. This is where data augmentation promises to overcome the lack of training data in the context of neural network engineering and classification tasks. The idea here is to reduce model over-fitting to the feature distribution of a small under-descriptive training data-set. We try to evaluate such data augmentation techniques to gather insights in the performance boost they provide for several convolutional neural networks on mel-spectrogram representations of audio data. We show the impact of data augmentation on the binary classification task of surgical mask detection in samples of human voice (ComParE Challenge 2020). Also we consider four varying architectures to account for augmentation robustness. Results show that most of the baselines given by ComParE are outperformed.

Face mask recognition from audio: The MASC database and an overview on the mask challenge

“…Illium et al. [60] explore a method that tries using Mel-spectrograms as features representing audio, and then employ some data augmentation technique combined with a CNN in order to solve the classification task at hand. Many augmentation techniques and CNN architectures are explored, and the best combination is used.…”

“… [38] 71.8 16.5 40.0 Deep Spectrum , BoWA, ComParE , auDeep 19 Illium et al. [60] 71.5 34.7 22.3 CNNs with multiple augmentations …”

“…In fact, the approaches by Illium et al. [60] and by [55] are the only approaches that have results worse than the baseline on the Test set, and both of them do not use ensembles. Only [63] managed to get good results without the use of ensembles.…”

“…This led them to attempt data generation using different approaches. A common approach for this is employing data augmentation techniques on the input data during training, like SpecAugment [51] , or other augmentations like time-shifting and adding Gaussian noise to the input data; this is employed in [54] , [60] . The most novel idea in this regard is introduced by Ristea and Ionescu [59] , who trained a GAN that generates data similar to the training data; this approach surpassed traditional techniques like SpecAugment [51] which masks frequencies and time, as well as image warping.…”

Identifying surgical-mask speech using deep neural networks on low-level aggregation

“…Note that our strategy takes a much less complicated pipeline for feature extraction compared with Ref. [18][19][20]. The contributions of this paper can be summarised as making use of low-level aggregation to achieve data augmentation and avoiding using external pre-trained models.…”