Approach
The seen or imagined images are recorded using fMRI pattern construction and an optional Generative Adversarial Networks (GAN) used to reconstruct the images from the predicted pattern. A CGAN (Conditional GAN), an improved version of GAN is presented along-with a Classifier, which improves the output of the whole system, thus obtaining a better output resemblance with the original image. The classifier is a pre-trained model of deep neural network, and the generated images mirror the stimulated images (both natural images and artificial shapes). The method successfully generalized the reconstruction to artificial shapes and natural images, indicating the model indeed ‘reconstructs’ or ‘generates’ images, and not simply throws an arbitrary result.
Research is on-going for this project :)
Gallery
(Click on image to view)
View Paper
Team Contact: Suyfan Parkar, Chaitanya Dandekar, Meet Dhanki