Underwater Distortion Target Recognition Network (UDTRNet) via Enhanced Image Features.

Area of Science:

• Robotics
• Computer Vision
• Oceanography

Background:

• Autonomous underwater vehicles (AUVs) struggle with target recognition due to limited labeled data and complex underwater environments.
• Light refraction and environmental complexity hinder AUVs from extracting complete target features.

Purpose of the Study:

• To propose an underwater distortion target recognition network (UDTRNet) to enhance image features for AUVs.
• To address the challenges of target recognition in data-scarce and visually complex underwater settings.

Main Methods:

• Utilized info noise contrastive estimation (InfoNCE) loss to extract significant image features.
• Constructed a dynamic correlation matrix to capture and extract spatial semantic features of underwater targets.
• Fused significant and spatial semantic features for model training using cross-entropy loss.

Main Results:

• The proposed UDTRNet demonstrated improved feature enhancement for underwater images.
• Achieved a 1.52% increase in mean average precision (mAP) for recognizing underwater blurred images.

Conclusions:

• The UDTRNet effectively enhances image features, leading to improved target recognition for AUVs in challenging underwater environments.
• The integration of InfoNCE loss and dynamic correlation matrix offers a robust solution for underwater object detection.