Transmission map-guided deep unfolding network for underwater image enhancement

Baocai Chang, Peng Duan, Jinjiang Li

Abstract: In recent years, with the development of maritime industries, the importance of underwater image enhancement and restoration has become increasingly prominent. However, underwater images often suffer from color distortion and low contrast due to the differential attenuation and absorption of light at different wavelengths by the medium. Most existing networks, however, adopt an end-to-end mapping approach that neglects prior information in the image enhancement process, resulting in a lack of interpretability. To address these challenges, we propose a transmission map-guided deep unfolded network for underwater image enhancement. Our method consists of three core components: Adaptive Mask Illumination Dynamic Prior (AMIDP), Transmission-Guided Multi-Scale Convolutional Dictionary (TGMCD), and Constant Spatial Aggregation Module (CSAM). AMIDP extracts the illumination characteristics of the image through a mask autoencoder and dynamic convolution, enabling the separate modeling of illumination and reflection information as shared and unique features, respectively. These features are then input into the TGMCD module, which is guided by the transmission map for iterative optimization. In this process, we replace the traditional proximal operator with a learnable multi-scale residual module, incorporating prior information and constraints to enhance model performance. Additionally, the CSAM is designed to strengthen information fusion across features, ensuring that the final enhanced image corrects distortions while retaining key details. Extensive experiments on multiple underwater datasets demonstrate that our method achieves state-of-the-art performance, validating its effectiveness and superiority. Our code is available at https://github.com/makabala/TGDU-master.

External IDs:dblp:journals/tjs/ChangDL25