Underwater Seafood Detection Using Deep Learning and Data Augmentation

Accurate seafood detection underwater is still a real challenge for modern fisheries, especially when it comes to resource monitoring or automated harvesting. In real-world conditions, underwater images are often blurry, discolored, or cluttered with background noise, all of which make feature extraction harder. On top of that,many marine species have irregular, elongated body shapes, which adds another layer of difficulty for reliable detection. To tackle these problems, we developed an underwater seafood detection model called Seafood Detection Deep Learning (SDDL). It builds on the YOLOv8 framework but integrates dynamic snake convolution to better handle elongated features like sea urchin spines or sea cucumber tentacles. This change lets the network deal more effectively with anisotropic shapes.Before training, we apply Contrast Limited Adaptive Histogram Equalization to boost image contrast and bring out finer details. We also use Focal Loss during training to reduce the impact of class imbalance. The dataset we used contains 5543 underwater images, which provides a realistic and fairly challenging testbed. SDDL achieves 84.22% mAP50 and 48.74% mAP50–95 on this dataset. When