Frequency-Adaptive and Multi-Scale Attention Network for Underwater Hull Fouling Segmentation

Underwater hull fouling segmentation is challenging because underwater images often suffer from:

Low contrast
Uneven illumination
Suspended particles and noise
Large variations in fouling scales
Strong similarity between fouling regions and background areas
To address these challenges, we propose FMA-UNeXt, an improved UNeXt-based semantic segmentation framework.

The proposed network consists of three key modules:

FADC: Frequency-Adaptive Dilated Convolution for enhancing low-frequency structures and high-frequency details
EMCAD: Efficient Multi-scale Convolutional Attention Decoder for reconstructing fouling regions with large scale variations
EMA: Efficient Multi-Scale Attention for improving global-local feature fusion and large-region segmentation

Full Changelog: https://github.com/lin2000921/fma-unext/commits/v1.0.0