Dual-Decoder RetinexNet: Underwater Image Enhancement via Adaptive Component Processing
TL;DR
The proposed dual-decoder RetinexNet architecture for underwater image enhancement based on illumination-reflectance decomposition incorporates several key innovations including Gaussian weight integration for mid-tone enhancement, Heaviside function-based adaptive control to prevent over-enhancement, non-local means denoising adapted for underwater texture patterns, and a weighted fusion strategy that dynamically balances Retinex reconstruction with individual component contributions.
Abstract
Underwater images suffer from wavelength-dependent attenuation, non-uniform illumination, and scattering effects that degrade visual quali-ty. This paper proposes a dual-decoder RetinexNet architecture for underwater image enhancement based on illumination-reflectance decom-position. The network employs two specialized decoders: an illumination decoder with large convolutional kernels (7×7, 9×9) to capture low-frequency lighting patterns, and a reflectance decoder with small kernels (3×3, 5×5) and spatial attention for high-frequency detail preservation. The method implements a six-step processing pipeline: (1) RetinexNet decomposition with underwater-specific loss functions, (2) illumination enhancement via guided filtering and adaptive thresholding, (3) reflectance processing through soft thresholding and aniso-tropic diffusion, (4) HSI color space transformation, (5) coordinated component refinement, and (6) adaptive weighted fusion. The fusion strategy combines components using spatially-varying weights α(x,y), β(x,y), γ(x,y) based on local decomposition and detail quality metrics. The method incorporates several key innovations including Gaussian weight integration for mid-tone enhancement, Heaviside function-based adaptive control to prevent over-enhancement, non-local means denoising adapted for underwater texture patterns, and a weighted fusion strategy that dynamically balances Retinex reconstruction with individual component contributions. The experimental results on underwater datasets show superior results compared to existing methods.