Adaptive fusion based deep learning framework for restoring underwater image quality using multi scale attention features

Images captured underwater often encounter quality degradation issues, such as blurring of details, low contrast, non-uniform illumination, and colour deviations. Like a significant challenge in computer vision (CV) and image processing, the development and restoration of underwater images are essential for many practical applications. Over the past few decades, underwater image restoration (UIR) has drawn a growing number of research efforts. Conventional image restoration technology typically begins with the image degradation method, identifies suitable image formers, and designs an intelligent optimizer model to produce the results. In recent times, deep learning (DL) has made considerable progress in the low-level vision domain. A DL is a data-centric model and may successfully fit the composite mapping procedure. It is appropriate for UIR, which is a composite issue with numerous degradations. In this manuscript, an Efficient Restoration of Underwater Images Using Multi-Scale Attention Features (ERUI-MSAF) model is proposed. The paper aims to develop an effective method for restoring underwater images by improving visibility and increasing the overall quality of the images. Initially, the adaptive bilateral filtering (ABF) technique is employed for pre-processing to enhance image quality by reducing noise. For image restoration, the ERUI-MSAF model integrates channel and spatial attention features to emphasize informative features and regions in underwater images adaptively. Finally, a fusion of DL is employed, using deep wavenet (DWN) for spatial attention features and EfficientNet for channel features, ensuring high performance and computational efficiency. The efficacy of the ERUI-MSAF methodology is examined under the EUVP and UIEB datasets. The comparison study of the ERUI-MSAF methodology revealed superior PSNR values of 34.258 and 29.0073 compared to existing models under dual datasets.