Attention-driven color balance and fusion algorithm in category-specific underwater image and video enhancement

Underwater imagery is critical for marine biology, oceanography, and underwater exploration, yet it suffers from challenges like color casts, low contrast, and haze due to light absorption and scattering. This paper proposes an innovative attention-driven color balance and fusion algorithm for category-specific underwater image and video enhancement, implemented as a standalone application. The method addresses diverse underwater conditions such as bluish hazy, greenish hazy, low-light bluish, and low-light greenish by classifying scenes based on dominant color and brightness. It integrates a dark channel prior (DCP) for dehazing, category-adaptive white balance, and attention-guided multi-scale fusion to restore clarity, color fidelity, and detail. The attention mechanism prioritizes salient regions, such as marine life or divers, ensuring enhanced visibility of critical features. The pipeline includes preprocessing, dehazing, color compensation, gamma correction, sharpening, and attention-guided fusion using Gaussian and Laplacian pyramids. Evaluated on Underwater Image Enhancement Benchmark Dataset (UIEB) and Enhancing Underwater Visual Perception (EUVP) datasets and real-time video from the Underwater Object Tracking (UOT32) dataset, the algorithm achieves good performance with average PSNR value ranging from 45dB, average FSIM scores of 0.99, and UIQM scores from 0.8276 to 1.0197. A user-friendly Graphical User Interface (GUI) application was also developed based on the technique that enables seamless image and video processing, displaying enhanced outputs alongside metrics. This approach offers a robust, real-time solution for underwater visual enhancement, adaptable to varied aquatic environments, with applications in marine research, archaeology, and filmmaking. The standalone application ensures practical deployment in real-world scenarios, advancing underwater imaging technology. The program and application will be available at https://github.com/bosadiq.