FSD-Net: underwater object detection based on frequency and spatial domain feature enhancement

Area of Science:

• Computer Vision
• Robotics
• Marine Technology

Background:

• Underwater visual conditions degrade object detection performance, leading to missed and false detections.
• Reliable autonomous underwater exploration is hindered by limitations in current object detection models.

Purpose of the Study:

• To address performance limitations in underwater object detection.
• To propose a novel detection model, FSD-Net, for complex underwater environments.

Main Methods:

• FSD-Net incorporates a Frequency Attention Convolution Module for feature preservation.
• A Multi-dimensional Feature Enhancement Module is used for semantic fusion to reduce false detections.
• Experiments involved ablation studies and comparisons with state-of-the-art methods on UTDAC2020 and Brackish datasets.

Main Results:

• FSD-Net achieved state-of-the-art performance on both UTDAC2020 (85.7% AP50) and Brackish (98.1% AP50) datasets.
• The model demonstrated improvements of 3.8% and 3.9% AP50 over baseline models on the respective datasets.
• Ablation studies confirmed the effectiveness of the Frequency Attention Convolution Module and Multi-dimensional Feature Enhancement Module.

Conclusions:

• FSD-Net's frequency-spatial enhancement effectively tackles underwater image degradation for robust autonomous exploration.
• The dual-module design provides a practical approach for optimizing detection models in challenging visual conditions.
• Future research will focus on developing lightweight versions of the FSD-Net model.