Generating Photorealistic Images from Human-Generated Sketches: A GAN-based Synthesis Approach for Enhanced Visual Realism
Mehedi Hasan, Abu Hasnayen Zillanee, Md. Golam Rabiul Alam · 2024
In the realm of advancing AI, the creation of lifelike images from sketches holds significant intrigue. Our proposed methodology employs Generative Adversarial Networks (GANs) to craft plausible images across various common categories. This system caters to a wide spectrum, encompassing individuals, animals, objects, and food items. Operating on the input of a sketch, our system employs a robust neural engine to scrutinize and produce new images, closely resembling reality. Additionally, we've implemented a data augmentation technique, significantly amplifying the diversity of available data for model training. Our proposed model has achieved an approximate accuracy of 96.36% in generating realistic images from sketches depicting people, and an accuracy of 40.63% for objects and animals. Furthermore, it achieved an average accuracy of about 76.63% in transforming stroke-based sketches into visual representations of people.