CPFU-Net: A Channel Position Wide Focus Mechanism based on U-Net for Building Image Segmentation Algorithm

Cheng Liu, Yue Pang, Tiezhu Li, Jiawei Sun, Xianyu Zuo, Lanxue Dang, Liming Zhou, Baojun Qiao

Abstract: Building image segmentation is widely used in building area estimation and urban construction planning. However, there are problems in the building segmentation in high-resolution remote sensing images, such as small targets misclassified easily, and edge details lost easily. To address the front issues, an improvement channel -position-focus-U-Net (CPFU-Net) model is proposed. A wide focus module (WFM) is designed to increase the receptive field which can achieve higher edge integrity of building segmentation through preserving edge details and focusing on the area of interest. The Squeeze-and-Excitation module is integrated into the channel attention convolution module (CACM) which improves the sensitivity of the model to channel information features. The CPFU-Net uses attention mechanisms and wide focus module (WFM) to improve skip connections which can fully utilize low-level and high-level semantic information to extract more accurate building features. Large numbers of experiments are implemented on both Satellite datasets and Massachusetts Building datasets to verify the effectiveness of CPFU-Net. The experimental results show that CPFU-Net reduces the probability of small target misclassification and preserve more edge details in building segmentation which improves the accuracy of building segmentation and achieves higher performance in OA, Recall, IoU, and F1-Score metrics.