In The Name of God
Master Thesis Defense Session
Computer Engineering, Artificial Intelligence Engineering
Supervisor:
Dr. Hossein Mahvash Mohammadi
Dr.Amir Hassan Monadjemi
Internal Reviewer:
Dr. Hamid Reza Baradaran Kashani
External Reviewer:
Dr. Alireza Karimian
Researcher:
Shamim Golafshan
Date: : 21 September 2022
Time: : 4:00 PM
Location:
Ansari building, Third floor, Dr. Braani Hall
Online link: lms.ui.ac.ir
Guest Account:
Username: computer
Password: computer1305
Topic:
Improving Quality of Reconstructed Seen Images Based on fMRI
Mind reading has been always one of the dreams of humans. Developing Artificial Intelligece and neuroimaging technics help people to achieve this goal. In science, this dreams called “Brain Decoding”. Brain decoding increases the awareness of the amount and mode of activity in different brain areas, which can be useful to know how the brain works and also identifies the nature of brain disorders and their improvement. Brain decoding defines as recording the signals generated in the brain during a person's activities (seeing a picture, film or dream), reconstructing and analyzing this information by artificial intelligence methods. If a person(subject) is watching an image, the decoding problem converts to the problem of reconstructing seen image. In this research, the quality of image reconstruction increased by using fMRI signals and artificial intelligence algorithms. For this purpose, in previous researches, fMRI signals were recorded during subjects were watching a number of color images (stimuli). These signals are divided into two categories: low-level visual areas and high-level visual areas. In addition, the stimuli were pre- processed to be more simple, and then their features are extracted by using a convolutional auto-encoder network. The low-level’s signals with the pre-processed images determine the outline of the shape of the stimulus by using a network called shape decoder. High-level's signals extract the meaning of the seen image by using a network called a semantic decoder. Finally, an generative adversarial network aggregates the output of these two models so that the reconstructed images are close to the stimulus images. In the previous researches that were accompanied by data augmentation, the structural similarity index was equal to 65.3, and in this research, without data augmentation, only by using a convolutional auto-encoder, the structural similarity index was equal to 68.4. Therefore, the quality and similarity of reconstructed images have been increased by using a convolutional Auto-encoder. As a result, by using these models, the quality of reconstruction can be increased and people can closer to realizing the dream of reconstructing people's thoughts.