SPPFU-Net: A Fusion Downsampling Network Focusing on Multiscale Features

Abstract

Semantic segmentation is a pixel-level classification. As a classic image segmentation network, U-Net is a common strategy for image segmentation. However, the traditional U-Net structure has limitations such as limited receptive field, poor modeling ability, and limited feature extraction ability. To deal with these problems, we proposed an improved U-Net with SPP attention and fusion downsampling (SPPFU-Net). A new multi-scale attention mechanism is added to strengthen features in different scales to learn richer semantic information, and a fusion downsampling module is designed to avoid the information loss caused by the downsampling. It shows good robustness to boundary information in the experiment. On the medical segmentation ACDC dataset, our model achieved good performance on DICE, HD95, and ASD coefficients metrics of 88.97%, 1.4303, and 0.3584, respectively, which are 0.41%, 0.85, and 0.24 higher than the classical U-Net. This demonstrates that the SPPFU-Net structure proposed in this paper can increase image segmentation accuracy.