DPMFformer: an underwater image enhancement network based on deep pooling and multi-scale fusion transformer

Due to light absorption and scattering, underwater images often suffer from color distortion, low contrast, and blurred details, seriously affects the effectiveness of advanced computer vision tasks. To address these degradation issues, this paper proposes an innovative underwater image enhancement algorithm, Deep Pooling and Multi-Scale Fusion Transformer (DPMFformer). The algorithm is composed of four key modules: the Dual-Balanced Multiscale Fusion Module (DBMF), the Deep Pooling Self-Attention Transformer (DPST), the Wavelet Sampling (WS), and the Global Spatial Feature Self-Attention Transformer (GSFAT). The DBMF module employs trainable color modules to simulate the grey-scale world theory, achieving inter-channel color balance. The DPST module enhances the network’s ability to extract information from feature regions through a deep-pooling layer and spatial attention mechanism. The WS module utilizes Harr wavelet sampling instead of conventional up- and down-sampling, preserving low-frequency information while improving the up-sampling outcome. The GSFAT module combines Swin Transformer (SwinT) and Position Embedding Cascading Transformer (PCET), enhancing the extraction of global information through position embedding and a sliding window self-attention mechanism, thereby improving the attention on the degraded regions of the image. Experimental results show that the proposed DPMFfomer is superior to existing underwater image enhancement methods.