AutoEnhancer: Transformer on U-Net Architecture search for Underwater Image Enhancement

Yi Tang, Takafumi Iwaguchi, Hiroshi Kawasaki, Ryusuke Sagawa, Ryo Furukawa; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 1403-1420

Deep neural architecture plays an important role in underwater image enhancement in recent years. Although most approaches have successfully introduced different structures (e.g., U-Net, generative adversarial network (GAN) and attention mechanisms) and designed individual neural networks for this task, these networks usually rely on the designer's knowledge, experience and intensive trials for validation. In this paper, we employ Neural Architecture Search (NAS) to automatically search the optimal U-Net architecture for underwater image enhancement, so that we can easily obtain an effective and lightweight deep network. Besides, to enhance the representation capability of the neural network, we propose a new search space including diverse operators, which is not limited to common operators, such as convolution or identity, but also transformers in our search space. Further, we apply the NAS mechanism to the transformer and propose a selectable transformer structure. In our transformer, the multi-head self-attention module is regarded as an optional unit and different self-attention modules can be used to replace the original one, thus deriving different transformer structures. This modification is able to further expand the search space and boost the learning capability of the deep model. The experiments on widely used underwater datasets are conducted to show the effectiveness of the proposed method.