Documents
Supplementary Material
Frequency Modulated Deformable Transformer for Underwater Image Enhancement
- Citation Author(s):
- Submitted by:
- Ashutosh Kulkarni
- Last updated:
- 7 February 2024 - 12:58pm
- Document Type:
- Supplementary Material
- Document Year:
- 2024
- Presenters:
- Adinath Dukre
- Paper Code:
- 2278
- Categories:
- Log in to post comments
Underwater images often suffer from degradation due to refraction, back-scattering, and absorption, resulting in color cast, blur, and limited visibility. Such degradation hampers higher-level computer vision applications in autonomous underwater vehicles. Existing methods for enhancing degraded images often fail to preserve fine edges and true colors. Hence, an effective pre-processing network is vital for underwater image enhancement. Addressing this need, we propose a frequency modulated deformable transformer network. Initially, our method extracts features using a multi-scale feature fusion feed-forward module. Further, a frequency modulated deformable attention module reconstructs fine-level textures in the restored image. We introduce a spatio-channel attentive offset extractor in the modulated deformable convolution for focusing on relevant contextual information. Additionally, we propose adaptive edge-preserving skip connections to propagate prominent edge features. Our method outperforms existing state-of-the-art methods, as demonstrated by comprehensive evaluations on synthetic and real-world datasets, and extensive ablation analysis. Testing code is provided in the supplementary material and will be made available upon final acceptance.