Prosthetic vision for the blind: Intelligent optimization of limited vision Testing trained computer agents for phosphene vision in a realistic environment

Karaca, berfu

Prosthetic vision for the blind: Intelligent optimization of limited vision Testing trained computer agents for phosphene vision in a realistic environment

Files

1062231_Karaca,B_MSc_thesis_2021.pdf (1.31 MB)

Authors

Karaca, berfu

Issue Date

2022-08-24

Language

en

Abstract

Blindness is a common societal problem that affects day-to-day functioning. Even though there is no effective treatment yet, there are some alternative ways such as neuroprosthetic visual implants. Although prosthetic vision does not provide normal vision, it does provide a rudimentary form of the environment through point-like flashes known as phosphenes which might still help basic activities like navigation. However, due to biological limitations, the current implants have low resolution. The limited capacity of resolution increases the need for optimal information extraction from the scene for efficient understanding of the environment. In line with this, there is a need for better image pre-processing techniques. One limitation of the studies testing phosphene vision is the necessary surgical operation for implants. For this reason, researchers found other techniques to test phosphene vision. One solution is testing sighted participants with wearable head-mounted displays (e.g., VR) that convert the real scene to processed phosphene vision. However, studies suggested that different image pre-processing techniques should be used for different contexts and scenes which requires an optimization adaptive to the scene. The variability in parameters to be tested and the need for optimization raise other challenges such as significantly increased number of tests for optimization problems which also means increased cost. This project will contribute to our knowledge on the potential use of trained Deep Reinforcement Learning agents to test the performance in particular tests such as navigation rather than human participants which is likely to minimize the cost and accelerate the process of optimization. Additionally, as these models will be used in real life, using a realistic virtual environment to test the behaviour of the trained agent and to optimize the parameters of simulated phosphene vision will provide more applicable results for future studies on prosthetic implants. Keywords: Phosphene vision, visual loss, visual prosthesis, deep learning, artificial intelligence, realistic virtual environment, computer agent