Multiscale Fusion Network for Rural Newly Constructed Building Detection in Unmanned Aerial Vehicle Imagery

Abstract

The accurate and rapid detection of buildings in rural areas is crucial in preventing illegal construction. The detection of newly constructed buildings can considerably reduce the cost associated with illegal construction management. However, because the number of newly constructed buildings is limited, unmanned aerial vehicle (UAV) images cannot be obtained from identical viewpoints and heights, inducing differences in the appearance and size of buildings. Thus, the detection of newly constructed buildings using multiform UAV images is the focus of this study. Herein, a multiscale fusion network is proposed to address the challenges associated with the diversity of UAV images and the limited number of newly constructed buildings. First, an adaptive weight channel attention network is used to optimize the building features obtained from UAV images. Then, the multiscale spatial pyramid network is used to realize feature fusion. Thereafter, a new dataset is created that includes 1590 images of various objects with high-quality annotation and a resolution of 1000 × 1000. Experimental results show that the proposed approach achieves an average detection speed of 23.18 frames per second, an accuracy of 72.8% for newly constructed buildings, an accuracy of 88.3% for completely constructed buildings, and an average precision of 80.6%. The accuracy of the proposed approach outperforms those of the baseline (increases of 40.9%, 2.1%, and 21.5%).