This fish does not exist: fish species image augmentation using stable diffusion

Machine learning provides a broad range of models for the detection and classification of marine fauna using underwater images. However, model performance remains largely dependent on the quality and quantity of the training, testing and validation imagery. Collecting these data is time-consuming, expensive and requires specialist knowledge. In some cases, it can be ethically questionable if data collection involves rare and endangered species. Augmentation aids in extending existing datasets by modifying the original images, and typically includes stretching, rotating, cropping and adding noise. In contrast to conventional augmentation approaches, we propose an additional augmentation by adding synthetically generated images. In this work, we show that stable diffusion, a generative model, has the potential to generate high-quality augmentation images of fish from multiple species. We provide a fish image augmentation image processing pipeline using open source imagery. Examples, limitations and suggestions for the generation of augmentation images are provided. Finally, we investigate the influence of different parameters for the embedding pipeline for the automated generation of imagery for fish species classification problems.