Transformer based Deep Learning Model for Underwater Image Enhancement

A parallel-guided underwater image improvement technique utilising a Transformer and CNN is suggested to rectify low contrast and colour distortion in underwater images. The approach employs a 3D position embedding model to furnish the Transformer with relative positional data, colour distortion metrics, and global characteristics of the feature map. A CNN encoder extracts local image features, while global features obtained from the Transformer are combined with the CNN-extracted local features via a feature modulation matrix. The CNN decoder enhances the image resolution, and the resulting feature map is input into a feature enhancement network, which produces the final output. The algorithm is trained using the current EUVP paired dataset. To validate the algorithm's superiority, underwater images exhibiting varied levels of colour distortion were chosen for qualitative comparison and quantitative analysis. The findings indicate that the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) of the enhanced underwater images surpass those of alternative comparison algorithms, with a notable enhancement in subjective quality, resulting in images characterised by more vibrant colours and greater clarity.