Artificial intelligence based semantic segmentation on aerial images with variational mode decomposition

Segmentation of objects in an image is a significant task in the domain of computer vision that entails algorithms customized to the nature of images and the type of their constituent objects. Although modern deep learning models apropos aerial images deliver state-of-the-art performance, major challenges in their segmentation of aerial images are fewer training samples, and segmentation of images with small and multiple objects. This paper proposes Variational Mode Decomposition (VMD) based semantic segmentation on aerial images for six different classes of objects. VMD extracts the predominant frequency components in an image, offering a better representation of the input image before segmentation. These frequency components are coupled with basic deeplearning models, ensuing in better segmentation results. This paper suggests a novel mode selection method for extracting coarser information aiding in efficient segmentation. Also, the paper proposes a solution for the problem of fewer training samples through VMD. The proposed methodology was experimented on images of three publicly available datasets: Dubai, Northwestern Polytechnical University-Very High Resolution (NWPUVHR10), and Aerial Image Datasets (AID). Based on the quantitative and qualitative analysis of the results on the Dubai dataset, it is inferred that the performance of basic deep neural networks has improved when coupled with the VMD technique. Among the deep neural networks experimented on in this paper, VMD-based UNet with Vision Transformer outperformed the other models in terms of visual comparison, mean Intersection over Union (mIoU), and mean F1 score.