Stable and Efficient Vision Model for AUV Perception via Multi-Task Learning with Gradient Normalization and Knowledge Distillation

Abstract

Computer vision in underwater environments faces significant challenges due to severe visual degradation caused by light absorption and scattering. These optical phenomena hinder semantic segmentation and depth estimation, which are critical for Autonomous Underwater Vehicles (AUVs) to perceive their surroundings, identify obstacles such as reefs, and execute safe motion planning. The primary objective of this study is to develop a highly stable and efficient Multi-Task Learning (MTL) architecture that simultaneously performs pixel-wise segmentation and dense depth estimation in real time. Regarding materials, the study uses the SUIM dataset, comprising 1,635 RGB images, in which depth ground truth was synthetically generated with the Depth Anything Model v2 to address the scarcity of paired underwater depth data. The proposed method features an EfficientNet-B7 encoder within a U-Net framework, incorporating Modified Gradient Normalization (MGN) to adaptively balance task gradients, Deep Supervision (DS) to enhance feature extraction, and Knowledge Distillation (KD) to refine depth predictions. The results indicate superior performance, with the model achieving a segmentation Intersection over Union (IoU) score of 0.745±0.0008 and a depth Mean Absolute Error (MAE) of 0.0797±0.0017, significantly outperforming the SegNet and DeepLabV3+ baselines while maintaining a fast inference time of 0.0116 seconds. The implication of this research confirms that balancing task gradients and supervising intermediate features are essential for robust MTL in degraded environments. This offers a viable solution for resource-constrained AUVs and sets a foundation for future integration with transformer-based attention mechanisms.