Semantic Segmentation of Spectral Images: a Comparative Study using FCN8s and U-NET on RIT-18 Dataset

Fully convolutional networks (FCN) have been used in various semantic segmentation applications. The spectral images consist of several bands that carry information that is beyond the visible spectrum of a human eye. Semantic segmentation has great potential in analyzing spectral images. Applying deep learning models on spectral images is limited due to the unavailability of the benchmark dataset. Recent trends in deep learning suggest methods to make complex problems like semantic segmentation more accurate or faster. In our study, we have adopted two FCN architectures (FCN8s and U-Net) and conducted a comparative study on a benchmark multispectral image dataset (RIT-18 Dataset). Since FCN architectures are so diverse in terms of use and remodeling, it can be easily extended for training multi-channeled image data. From our study, FCN8s perform better than U-Net on the segmentation of the RIT-18 dataset.