Examining the Role of Perceptual Quality in Underwater Image Classification

Underwater image classification faces significant challenges due to poor visibility conditions, color distortion, and light scattering that degrade image quality. Building upon existing work in underwater computer vision and image enhancement, we investigate the impact of perceptual quality on binary fish classification performance using the EUVP dataset [1], which provides paired underwater images with good and poor perceptual quality. We evaluate five CNN architectures (ResNet50, MobileNetV2, DenseNet121, EfficientNetB0, and EfficientNetV2B3) under three training configurations: using poor-quality images, good-quality images, and a combined dataset. Our experimental results confirm and extend previous findings, showing that models trained on good-quality images generally achieve better performance when tested within the same quality domain compared to models trained and tested on poor-quality images. Training on combined datasets enhanced the models' ability to generalize across both test qualities simultaneously, as expected. When testing on poor-quality images, models trained on the combined dataset consistently demonstrated superior performance (with only one exception across all five CNN architectures). When testing on good-quality images, two of the five CNN architectures actually performed best when trained on the combined dataset, while the remaining three mostly showed only marginally better results when trained exclusively on good-quality data. The dataset and code are available at: https://github.com/davsalmeida/perceptual-quality-classification.