MSCT: Multi-Scale Conv-Transformer for Underwater Image Enhancement

In order to solve the problems of underwater image blurring, color deviation and low contrast, researchers have proposed various underwater image enhancement methods. Among them, the methods based on convolutional neural networks (CNNs) have difficulty extracting global information, while the methods based on transformer struggle to capture local information. To address the aforementioned issues, we propose a multi-scale Conv-Transformer (MSCT) model based on an encoder-decoder architecture, which introduces a novel self-attention mechanism and an upsampling method. Specifically, we utilize the multi-head convolution channel-attention (MCCA) in place of the self-attention, effectively capturing both global structures and detailed features. Simultaneously, we propose a multi-scale feature fusion module (MFFM) that incorporates local probability maps from the encoder’s output features into the upsampling process, enhancing the ability to restore detailed textures. Experimental results across several datasets demonstrate that the proposed method outperforms other underwater enhancement techniques in terms of various metrics.