Fusion of Traditional and Deep Learning Models for Underwater Image Quality Assessment

Underwater image quality is significantly affected by the unique challenges of the underwater environment, including light attenuation, color distortion, backscatter, and the presence of particulate matter. These factors reduce visibility, alter color perception, and introduce noise and artifacts that degrade the overall quality of captured images. Ensuring high-quality underwater images is essential for various applications, including marine biology, environmental monitoring, underwater navigation, and industrial operations. This research proposes a novel approach to evaluating underwater image quality by integrating conventional machine learning methods with advanced deep learning architectures. The fusion process, employing Support Vector Regression (SVR), known as FTDUIQA (Fusion of Traditional and Deep Learning Models for Underwater Image Quality Assessment), combines well-established deep learning-based models - CNNIQA (Convolutional Neural Network for No-Reference Image Quality Assessment), DBCNN (Deep Bilinear Convolutional Neural Network), and HyperIQA (HyperNet for Image Quality Assessment) - with the classical IQA model, MSAEQA (Multi-scale and Antagonistic Energy Distribution). Additionally, we incorporate a modified deep learning model, MAUIQA (Multi-scale AttentionNet for Underwater Image Quality Assessment), which has been adapted from its image super-resolution application. This hybrid approach leverages the strengths of both deep learning and traditional techniques to provide a robust and comprehensive assessment of underwater image quality.