Underwater Image Enhancement via Residual Dense Block and Feature Pyramid Attention

When light travels underwater, as a result of absorption and scattering, underwater images show color casts and low contrast. Recent years have witnessed a boom in approaches of underwater image enhancement which based on deep learning methods. However, the existing methods do not utilize the rich hierarchical features in the degraded underwater images. We propose a new underwater image enhancement approach to address these problems, which contains the residual dense block and the feature pyramid attention. The abundant hierarchical features in degraded underwater images are extracted through densely connected convolutional layers. Besides, the feature pyramid attention provides accurate pixel level attention for high-dimensional features of the network. Our proposed approach can remove color casts and improve image contrast effectively, outperforming other traditional and deep learning methods on a real-world dataset.