Multi-Variable Expanded Latent Space Autoencoder for Underwater Image Classification

We propose a multi-variate representation approach built with autoencoders, named MVELSA, for underwater image classification. Basically, the idea is that the system instantiates multiple independent autoencoders rather than relying on a single to feature extraction, which allows our system to capture different statistical nuances of the data in the latent space of each autoencoder. The evaluation is performed on the AQUA20 benchmark, which reflects realistic underwater conditions and highly asymmetric class distributions, in which we conducted a comparative study with consolidated deep learning architectures, such as ResNet-18 and YOLOv8n. Experimental results indicate that the convolutional models trained via transfer learning achieve competitive global accuracy but exhibit limited performance on minority classes, as reflected by lower macro-averaged recall and F1-score. In contrast, the MVELSA approach demonstrates balanced performance across classes, under several experimental conditions. Although performance gains are data-dependent, our findings suggest that combining different latent space representations is a promising alternative for underwater image classification, particularly in scenarios with scarce and imbalanced data.