U-TransCNN: A U-shape transformer-CNN fusion model for underwater image enhancement

Abstract

Underwater imaging faces significant challenges due to nonuniform optical absorption and scattering, resulting in visual quality issues like color distortion, contrast reduction, and image blurring. These factors hinder the accurate capture and clear depiction of underwater imagery. To address these complexities, we propose U-TransCNN, a U-shape Transformer- Convolutional Neural Networks (CNN) model, designed to enhance underwater images by integrating the strengths of CNNs and Transformers. The core of U-TransCNN is the Global-Detail Feature Synchronization Fusion Module. This innovative component enhances global color and contrast while meticulously preserving the intricate texture details, ensuring that both macroscopic and microscopic aspects of the image are enhanced in unison. Then we design the Multiscale Detail Fusion Block to aggregate a richer spectrum of feature information using a variety of convolution kernels. Furthermore, our optimization strategy is augmented with a joint loss function, adynamic approach allowing the model to assign varying weights to the loss associated with different pixel points, depending on their loss magnitude. Six experiments (including reference and non-reference) on three public underwater datasets confirm that U-TransCNN comprehensively surpasses other contemporary state-of-the-art deep learning algorithms, demonstrating marked improvement in visualization quality and quantization parameters of underwater images. Our code is available at https://github.com/GuoRuige/UTransCNN.