Asymmetric Attention Upsampling: Rethinking Upsampling For Biological Image Segmentation
Chunyu Dong, Qunfei Zhao, Kun Chen, Xiaolin Huang · 2021
Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China Rescaling a feature map could be a key issue of biological image segmentation. Nearly all the existing upsampling strategies concentrate only on local information and attempt to expand the reception field via deepening the network. In this paper, we present Asymmetric Attention Upsampling (AAU) for biological-image segmentation. AAU utilizes the information of low-level feature maps to rescale the high-level feature maps smartly through spatial pooling and attention mechanisms. It consists of two attention variants: Asymmetric Spatial Attention (ASA) and Asymmetric Channel Attention (ACA). The Asymmetric Attention Upsampling Network (AAU-Net) combines several AAU blocks to achieve better segmentation performance. Experiments on the Kvasir-SEG data set reveal the effectiveness of our work. AAU-Net outperforms other state-of-the-art methods for polyp segmentation while not consuming many resources.