MSFFT-Net: A multi-scale feature fusion transformer network for underwater image enhancement

Abstract

Due to light attenuation and scattering, underwater images typically experience various levels of degradation. This degradation adversely affect object detection and recognition in underwater imagery. Nevertheless, the methods based on convolutional networks have limitations in capturing long-distance dependencies and the methods based on generative adversarial networks exhibit a poor enhancement effect on local detail features. To address this issue, we propose a Multi-Scale Feature Fusion Transformer Network (MSFFT-Net). We design an Underwater Transformer Feature Extraction Module (UTFEM) for conducting window self-attention calculations via maskless reflection filling, thereby enabling the capture of long-distance dependencies. The Channel Transformer Selective Kernel Fusion module (CTSKF) is devised as a replacement for the skip connection. By employing one-stage multi-scale feature coding recombination and two-stage selective kernel (SK) fusion, the model enhances its focus on local detailed features. Extensive experiments on three publicly available datasets demonstrate that our MSFFT-Net achieves better performance than some well-recognized technologies.