Improving Building Extraction by Using Knowledge Distillation to Reduce the Impact of Label Noise

Abstract

Building extraction using deep learning techniques has advantages but relies on a large number of clean labeled samples to train the model. Complex appearance and tilt shots often cause many offsets between building labels and true locations, and these noises have a considerable impact on building extraction. This paper proposes a new knowledge distillation-based building extraction method to reduce the impact of noise on the model and maintain the generalization of the model. The method can maximize the generalizable knowledge of large-scale noisy samples and the accurate supervision of small-scale clean samples. The proposed method comprises two similar teacher and student networks, where the teacher network is trained by large-scale noisy samples and the student network is trained by small-scale clean samples and guided by the knowledge of the teacher network. Experimental results show that the student network can not only alleviate the influence of noise labels but also obtain the capability of building extraction without incorrect labels in the teacher network and improve the performance of building extraction.