Adaptive Deep Learning Network With Multi-Scale and Multi-Dimensional Features for Underwater Image Enhancement

Nianzu Qiao, Lu Dong, Changyin Sun

Abstract: Owing to the turbidity, light absorption, and scattering of water, the quality of the acquired underwater images is worse and seriously affects the processing of subsequent vision tasks. To solve this problem, a high-performance underwater image enhancement model is proposed in this paper. We design a convolutional structure and a pooling structure. The convolution structure is named ‘flower convolution’, which can well expand the perceptual field and retain the ability to extract local features. The pooling method is called ‘dewater pooling’, which not only considers the overall features but also highlights the salient features. In addition, we propose a novel attention mechanism and a multi-dimensional feature fusion method, which can well fuse high and low-dimensional features, strengthen salient features, and suppress unimportant features. Moreover, we design an adaptive loss function, which can evaluate the quality of underwater images more reasonably. At last, we train and test the model on a real-world underwater image/video dataset. The experimental results demonstrate that the model outperforms the state-of-the-art methods in both qualitative and quantitative evaluation. The experimental results of the ablation study demonstrate the usefulness of the various components. Further applied research also shows the superior capability of the model.