Radial basis function networks with lightweight multiscale fusion strategy-based underwater image enhancement

A novel underwater picture enhancement approach under non-uniform lighting is presented to solve the issues of underwater photographs with unevenness due to additional lighting in deep-sea and night-time environments. Water suspended particles can cause image noise, low contrast, and colour deviation. The heterogeneous feature fusion module aims to combine multiple levels and levels of features with improving the network's ability to perceive semantic and specific information. The capability of autonomous underwater and remotely driven cars to explore and comprehend their environments is contingent on improving underwater images, a crucial low-level computer vision challenge. Recent applications of deep learning models include enhancing aquatic image quality and resolving several computer vision problems. Although several deep learning-based techniques exist for enhancing underwater images, their implementation is challenging due to the high memory and model parameter requirements. We propose a solution based on radial basis function networks (RBFN) for lightweight multiscale data fusion (LMFS). The LMFS incorporates diverse branches with varying kernel sizes to generate multiscale feature maps. The proposed RBFN-LMFS The convolution layer with jump connection and the attention module produces the output from the feature extraction module, which aims to extract various features at the network's beginning. The outcomes of our experiments on diverse data sets demonstrate that our proposed RBFN-LMFS technique performs well in processing both synthetic and authentic underwater images and successfully recovers image colour and texture characteristics. The visual output is superior to existing underwater image enhancement algorithms and is consistent with aspects of human vision.