Temples Restoration using Gated Convolution and Contextual Attention in Generative Adversarial Networks
Vihar Devalla, Venkata Krishnarjun Vuppala, Srinivasa Raghavan S, Tejas Kumar S, Arti Arya · 2023
India in the past is renowned for its magnificent and extraordinary architecture. Due to several invasions and climatic changes over the years, these magnificent works of art did not last very long. A novel Deep Learning approach is proposed in this work to digitally restore broken and eroded pieces of heritage temples and sculptures in India. This approach uses Gated Convolution and Contextual Attention layers added to Generative Adversarial Networks(GANs). The learning of this model is enabled by using copyright-free images of Indian temples and sculptures after going through data collection and data pre-processing steps. The data pre-processing involves resizing and applying free-form masks on missing pieces. The final images restored by this model have an L1 loss of 0.29 after training the model for 7,10,000 iterations. The whole objective of this work is to rebuild images of ancient temples as they were before being damaged by several invasions and climatic changes, thereby digitally restoring the past glory of India in a small way.