Publication

Abstract : Underwater environments present unique challenges for image analysis due to light attenuation, scattering, and color distortion, which significantly degrade image quality. These facts make simple operations such as image captioning, which is crucial in marine sciences, search, and conservation, challenging. Current datasets such as Flickr8K contain mostly terrestrial images and do not possess features distinctive to underwater environments therefore direct training of models is not possible in this context. To cover this gap, the current work combines a Transformer-based image captioning model with an enhanced underwater-specific image augmentation set. These augmentations mimic real-life underwater consequences such as RGB attenuation, Gaussian blur, muddy particles, noise, and gradient-based illumination distortions. The implemented model, learned and recursively tuned on the Flickr8k dataset, proves the ability to generate reasonable captions for underwater images even though the dataset belongs to Flickr8k providing a terrestrial environment. This is proven by the enhanced BLEU scores as well as from the results of the experiment which trained the translation model in less than ten percent of the dataset to show its flexibility on underwater scenarios but without necessitating the use of an underwater dataset.