Attention-guided multi-path cross-CNN for underwater image super-resolution

Raw underwater images usually suffer from quality degradation, and their resolutions are lower. To obtain high-resolution underwater images, some super-resolution (SR) algorithms have achieved great visual effect based on the excellent ability of deep convolution neural network. However, most previous works fail to consider the full use of inner information. Besides, simply widening and deepening the network contribute little to performance improvement. In this paper, an attention-guided multi-path cross-convolution neural network (AMPCNet) is proposed for underwater image SR. We present a multi-path cross (MPC)-module, which contains residual blocks and dilated blocks, to enhance the model’s learning capacity and increase abstract feature representation. Specifically, the design of the cross-connections realizes the mutual fusion of local features learned by residual blocks and multi-scale features acquired by dilated blocks. And the combination of dilated convolution and ordinary convolution achieves a trade-off between performance and efficiency. Furthermore, an attention block is presented to adaptively rescale the channel-wise features for more discriminative representations. Finally, the upsample block helps the reconstruction of HR images. Experimental results demonstrate the superiority of our AMPCNet network in terms of both quantitative metrics and visual quality.