Real-time underwater object detection via frequency-domain dynamics and spatially enhanced feature modulation

Area of Science:

• Computer Vision
• Marine Engineering
• Deep Learning

Background:

• Underwater object detection faces challenges due to poor visibility, low contrast, and texture loss.
• Deploying deep learning models on resource-constrained platforms requires balancing accuracy and inference speed.

Purpose of the Study:

• To propose a novel, lightweight deep learning framework for efficient and accurate underwater object detection.
• To enhance feature extraction and mitigate detail loss in complex underwater environments.

Main Methods:

• Developed a FasterFDBlock backbone using Partial Convolution and Frequency-domain Dynamic Convolution for adaptive noise suppression and edge enhancement.
• Introduced an AIFI-SEFN encoder with a Spatially Enhanced Feed-Forward Network to integrate global and local features.
• Implemented a Multi-scale Feature Modulation (MFM) module for dynamic feature weighting to improve robustness against scale variations and interference.

Main Results:

• Achieved a mean Average Precision (mAP) of 72.1% on the UTDAC2020 dataset, outperforming the baseline by 1.7%.
• Reduced model parameters by 27.1% and GFLOPs by 24.6%.
• Attained an inference speed of 72.6 FPS, demonstrating high efficiency.

Conclusions:

• The proposed lightweight framework offers a significant improvement in accuracy and efficiency for real-time underwater object detection.
• The novel backbone, encoder, and modulation modules effectively address challenges in underwater imaging and model deployment constraints.