Stonefish and Non-Stonefish Dataset

A specially tailored dataset, YU-SF, was created. In total, 700 underwater images were obtained from publicly available sources, including Dreamstime, Saltwater, Kaggle, and social media posts shared by divers. To maintain uniformity and enhance the model’s generalization capability, various preprocessing techniques were applied to the dataset. To meet the input dimension requirements of pre-trained deep learning architectures, all images were standardized in size to 224x224 pixels. The images were saved in PNG format, which preserves quality through lossless compression, ensuring that the most critical texture and color information is retained. Data augmentation was performed using TensorFlow's ImageDataGenerator to mitigate overfitting and enhance the model's robustness by artificially introducing variability. The data augmentation included rotations by ±45°, horizontal flips with a 50% probability, width and height shifts of up to 20%, and a zoom factor of 0.1. Utilization of the nearest fill mode ensures spatial coherence. These alterations enriched and broadened the dataset’s quantity and diversity. The model's ability to correctly identify camouflaged stonefish across their different natural environments improved markedly as a result.