Research on Attention-Driven Unsupervised Underwater Image Enhancement

Images taken underwater often experience color discrepancies and diminished clarity due to the effects of light dispersion and absorption. Addressing this challenge, our study introduces an underwater image enhancement technique, VAUIR, leveraging an attention-driven approach. This technique is rooted in a refined Koschmieder model, segmenting each image into three distinct elements. By employing the attention mechanism, it minimizes noise interference and elevates image fidelity. The process involves treating original underwater photos with a scene brightness prediction network (J-Net), a transmission map prediction network (T-Net), and a universal background light estimation unit (GB). Subsequently, the images are reformed using the image segmentation model. The generated degraded image and the original underwater image are fed into J-Net and Feature Fusion Attention Network (A-Net) to get a clear image. In addition, we employ multi-layered supervisory mechanisms for supervising the model, namely: a reconstruction loss is employed to supervise the J-Net network to emphasize the ability of the potential components to reconstruct the input image; a consistency requirement is applied to guide the A-Net network, ensuring that both the initial underwater photo and the reprocessed image maintain analogous structural tiers. Additionally, a color correction loss is employed to rectify any color deviations in the restored image. Test outcomes indicate that the VAUIR technique delivers commendable results in restoring underwater images, showcasing efficient recovery quality and processing speed.