Transfer Learning Based Underwater Image Segmentation using DenseNet-201 with U-Net Architecture

Underwater image segmentation is crucial for marine conservation, environmental monitoring, and underwater robotics applications. However, the challenges posed by light attenuation, color distortion, and low visibility in underwater environments make segmentation tasks inherently difficult. This paper proposes a novel approach combining DenseNet-201 as the encoder backbone with U-Net architecture for multiclass underwater image segmentation. Our method leverages transfer learning from ImageNet pre-trained weights and employs advanced loss functions (Dice Loss and Categorical Focal Loss) to address class imbalance. 
Experimental results on the SUIM dataset demonstrate that our proposed DenseNet-201 based U-Net architecture achieves a mean Intersection over Union (IoU) of 77.77%, outperforming existing methods such as SUIM-NetRSB (75.75%) and DeepLab v3+ (72.88%). The model achieves an F1-score of 0.8485 on the training set and 0.7439 on the validation set, demonstrating superior generalization and robustness for underwater image segmentation tasks across 8 distinct classes. 
Keywords: Underwater Image Segmentation, Transfer Learning, DenseNet-201, U-Net, Deep Learning, Semantic Segmentation