SCAUIE-Net: Underwater Image Enhancement Method Based on Spatial and Channel Attention

Underwater image enhancement is a Low-Level Vision task that plays an important role in marine resource development, but the light absorption and scattering cause severe underwater image quality degradation. To solve these problems, this paper proposes a neural network based on a spatial and channel attention module that reinforces the network’s attention to channel and spatial information. The network’s Confidence Generator can precisely extract feature maps from multi-scale underwater images. Meanwhile, we propose a new training loss function by mixing perceptual, MS-SSIM and MAE loss functions to further improve the contrast in high-frequency, colors and luminance. For training, this paper also uses a feature fusion strategy: Firstly, augmenting the training underwater images by Gamma Correction, White Balance and Histogram Equalization algorithms to remove color cast, lighten up dark regions and improve the contrast. Then, fusing the enhancing images with confidence maps predicted from the Generator. The network was validated in the UIEB dataset and obtains efficient improvements on Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics, yielding a PSNR of 22.9286 and SSIM of 0.9290. Experimental results on real-world underwater images demonstrate that the proposed method performs well on different underwater scenes.