For the training of deep neural networks as well as for other machine learning methods, it can be very useful to increase the available data-set(s) by data augmentation, i.e., by generating modified copies of the data. In the context of Underwater Human Machine Interaction (U-HRI) [1], we have developed multiple methods for physically plausible data augmentation of underwater images.
More precisely, the methods actually degenerate the available images in ways that correspond to forms of image degradation that can be commonly found in underwater vision scenarios. The code for underwater image augmentation by physically plausible image degradation is released on Github as part of a larger software packet for underwater gesture recognition [2].
The image data augmentation was applied to the CADDY Underwater Gestures Dataset [3] when using different classical machine learning (ML) and deep learning (DL) methods for recognizing the gestures of divers [2]. But the methods are of course usable for any kind of applications of underwater vision where deep learning, respectively data augmentation are used.
References
[1] A. Birk, “A Survey of Underwater Human-Robot Interaction (U-HRI),” Current Robotics Reports, Springer Nature, vol. 3, pp. 199-211, 2022. https://doi.org/10.1007/s43154-022-00092-7 [Open Access]
[2] A. G. Chavez, A. Ranieri, D. Chiarella, and A. Birk, “Underwater Vision-Based Gesture Recognition: A Robustness Validation for Safe Human-Robot Interaction,” IEEE Robotics and Automation Magazine (RAM), vol. 28, pp. 67-78, 2021. https://doi.org/10.1109/MRA.2021.3075560 [Preprint PDF]
[3] A. G. Chavez, A. Ranieri, D. Chiarella, E. Zereik, A. Babic, and A. Birk, “CADDY Underwater Stereo-Vision Dataset for Human-Robot Interaction (HRI) in the Context of Diver Activities,” Journal of Marine Science and Engineering (JMSE), spec.iss. Underwater Imaging, vol. 7, 2019. https://doi.org/10.3390/jmse7010016 [Open Access]