Evaluating the Impact of Histogram Modification on Building Segmentation using Deep Learning

Abstract

Despite UAV-photogrammetry being a potential solution for data acquisition, the mapping process cannot always be conducted automatically due to various factors. Therefore, it is imperative to develop automation methods that can be effectively applied to detailed-scale mapping. This study investigates the effectiveness of utilizing airborne imageries as training datasets for building segmentation in UAV imagery using deep learning. In addition to segmenting the original UAV imagery, we enhanced accuracy by employing histogram modification. The results of our study demonstrated that this approach effectively improved the quality of building segmentation results obtained from UAV imagery. Initially, the original UAV imagery yielded an accuracy score of 69.51% and an IoU score of 71.50%. However, after implementing histogram modification, the accuracy score increased to 73.18%, while the IoU score experienced a more significant improvement, reaching 80.52%. These findings indicate that improving deep learning results is not only about increasing training data or refining model design; modifying test data also plays an important role in achieving better performance.