Bio-Inspired Networks;: From DoG to CNN.
Date of Submission
December 2018
Date of Award
Winter 12-12-2019
Institute Name (Publisher)
Indian Statistical Institute
Document Type
Master's Dissertation
Degree Name
Master of Technology
Subject Name
Computer Science
Department
Machine Intelligence Unit (MIU-Kolkata)
Supervisor
Ghosh, Kuntal (MIU-Kolkata; ISI)
Abstract (Summary of the Work)
There has been a quest to understand the structure and functionality of the brain, in general, and the human visual system, in particular, since centuries. One reason that the quest prevails among engineers too, is because of the belief that if we can understand the human brain and visual system, it could help us design and develop models for similar tasks with similar high performance and accuracy. This, on the other hand, may perhaps also aid to arrive at a unified computational model for vision and may open new avenues in artificial vision. The persual of the quest has thus resulted in a good amount of research in the field of bio-inspired models, especially inspired from mammalian vision system. In this work, a study of several computational models for different levels of vision has been performed and its application in varying domains have been explored.The models, studied here, revolve around the central theme of David Marr’s organisation of hierarchy of vision as an information processing system and the bio-inspired models for each level of it. A major part of the work revolves around mid-level visual representation of an image and study of biologically inspired models like difference of Gaussian, which is also a computational model for the response of retinal ganglion cells, and those of Lateral Geniculate Nucleus (LGN). The basic DoG along with its variants like Extended Difference of Gaussian (EDoG), Oriented Difference of Gaussian (ODoG) and Dynamic EDoG have also been explored as possible unified approaches to low-level and mid-level vision.For instance, the EDoG is already known for modeling of the response of the non-Classical Receptive Field (nCRF) of retinal ganglion cells with an extra Gaussian in comparison to DoG that leads to better edge map(Ghosh, Sarkar, and Bhaumik, 2005b). This dissertation work consist of three approaches to explore bio-inspired models for the three levels of Marr’s hierarchy for vision. The first part of the work is dedicated to using the EDoG for understanding how geometry around an object effects its perceived size. In this part, structural and geometric information present in an image is used to find out how size depends on shape using the EDoGmodel. To do this a geometric illusion, namely The Muller Lyer Illusion (MLI) has been used for study. More precisely, EDoG has been used to understand how the lengths of the lines in the illusion are perceived. Further, the role played by geometry of and around an object in perceiving its length is observed. To do this, the relation between a critical parameter of the illusion (angle between the wings) and the induced illusion is also investigated. The results obtained from computational model have been compared with the experimental results for verification. This shows that the EDoG can be a plausible model of mid-level vision, beyond edge representation.The second part of this work is devoted to study a modified adaptive version of EDoG model (Wei, Wang, and Lai, 2012). By using the EDoG model with reverse control mechanism of vision, the brightness intensity information contained in an image has been used to give a good mid-level representation.
Control Number
ISI-DISS-2018-382
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
DOI
http://dspace.isical.ac.in:8080/jspui/handle/10263/6948
Recommended Citation
Ojha, Rahul Kumar, "Bio-Inspired Networks;: From DoG to CNN." (2019). Master’s Dissertations. 255.
https://digitalcommons.isical.ac.in/masters-dissertations/255
Comments
ProQuest Collection ID: http://gateway.proquest.com/openurl?url_ver=Z39.88-2004&rft_val_fmt=info:ofi/fmt:kev:mtx:dissertation&res_dat=xri:pqm&rft_dat=xri:pqdiss:28843279