PECF-Net: Multi-path underwater image enhancement network via multi-visual dimension features

Abstract Due to the complex scattering and absorption effects in multifactorial underwater conditions, the quality of original underwater images is often significantly degraded, which hinders the development of marine resource engineering. Therefore, this paper proposes (PECF-Net), an underwater image enhancement network based on multi-visual input to achieve multi-feature interactive fusion. This network primarily obtains multi-dimensional feature images through preprocessing and generates refined multi-dimensional feature outputs via the Neuron Difference Layer (NDL). By leveraging the differences among channel neurons, it adjusts the weight ratio of pixel values across different ranges, thereby achieving an enhanced color feature output while preserving key local information. Additionally, a channel-space parallel attention mechanism is developed to calibrate the detailed features of the image and subsequently facilitate the effective interactive fusion. A comprehensive evaluation of the model using public datasets, such as LSUI, demonstrates its excellent learning capability. The experimental results indicate that the model outperforms state-of-the-art methods by 8.8% in terms of the PSNR index, based on testing a randomly selected subset of the LSUI dataset. Furthermore, the effectiveness of each module in the proposed PECF-Net is validated through comparison with existing methods.