Shadow Detection using GAN with Residual CNN Discriminator
Sandeep Kumar Yadav, Satyajeet Singh, Antoreep Jana, Seba Susan · 2020
The objective of this paper is shadow detection. For this purpose, we have proposed a Generative Adversarial Network (GAN) model with a unique generator-discriminator combination. The U-Net convolutional neural network architecture comprising of an encoder and decoder, that is popularly used for image segmentation, is used as the generator. The discriminator module is based on the ResNet pre-trained network architecture, and uses residual convolutional blocks. The residual blocks considerably improve the performance of our network. Our model was evaluated on the SBU shadow detection dataset and the results prove the efficacy of our method as compared to the state of the art.