Automatic Aircraft Shadow Removal from Remote Sensing Images using Mask-ShadowGAN

Abstract

The topic of the project is Automatic Aircraft Shadow Removal from Remote Sensing Images using Mask-ShadowGAN. The objective of this research is to introduce a new method to remove the shadow in airplane images. We use remote sensing images in this work. The new procedure learns to remove the shadow automatically by using shadow and shadow-free images. Unlike other algorithms, Mask-ShadowGAN requires no identical training image set. This aspect eases the data collection process for the user. The framework develops in a Python environment using PyTorch. We can see that the generated shadow-free images have shadows thinned or faded out, but the airplane shape is still intact. We then evaluate the results using Root-Mean-Square-Error with generated shadow-free images and Jaccard Index with their binary images. The binary images are obtained through a custom image processing technique. Our score is 0.7799, while the Modified DSS, which is traditional CNN with post-processing and L2- Normalization, the score is 0.7775, which the Mask-ShadowGAN performs better than the Modified DSS.