Adaptive Channel Attention-Based Deformable Generative Adversarial Network for Underwater Image Enhancement

In this paper, to effectively strengthen quality of underwater image enhancement from both channel and spatial viewpoints, an adaptive channel attention-based deformable generative adversarial networks (ACADGAN) framework is established. Main contributions are as follows. 1) By virtue of multi-branch convolution architecture with dilated convolution mechanism, the adaptive channel attention (ACA) is devised, such that channel weight can be adaptively recalibrated, and thereby significantly contributing to preserving content features from channel viewpoint. 2) By augmenting offset position of sampling point with respect to convolution kernel, the deformable convolution network (DCN) is created, such that detailed information of underwater image can be dramatically retained from spatial aspect. 3) The ACADGAN scheme is eventually proposed by integrating ACA and DCN modules with a deep generative adversarial network. Comprehensive experiments demonstrate the remarkable effectiveness and superiority of the developed ACADGAN scheme.