Tackling data scarcity in sonar image classification with hybrid scattering neural networks

Data scarcity remains the main challenge when developing deep learning models for sonar image analysis. Although dataset augmentation with synthetically generated images has been proposed, these methods are far from optimal as they are unable to capture the range of physical factors affecting sonar images, given the small data regimes used for their training. This work focuses on an alternative solution and investigates the learning of suitable representations for classifying small-sized sonar datasets. To achieve this, we propose a new approach that entails the combination of convolutional and scattering neural networks, a wavelet-based neural network that produces feature map representations robust to image variations. Our experiments show that these representations are easier to classify, leading to a performance increase of 4.5 percentage points in F1-score for the combined network compared to a plain convolutional neural network. Furthermore, we interpret the representation obtained by the scattering transformation as robust feature descriptors, where the geometric shapes of underwater objects are rendered prominent and stable to minor sonar distortions.