Our project develops a multi-label image classification system to address the poor optical conditions encountered during underwater visual inspections. Manual inspections are dangerous and costly, and are further compromised by turbidity, blur, low light, and sensor noise. To mitigate these risks, we trained and benchmarked four deep learning classification models for use on autonomous underwater vehicles, complemented by interpretability features including Grad-CAM heatmaps. We address domain generalization by training models on artificially degraded data, producing a system capable of reliably labeling anomalies under both ideal ("clean") and degraded ("blurred") conditions.