Evaluating prosthetic vision by decoding phosphene images
Globally, over 30 million people su↵er from blindness. Visual prosthetics can help such people, by stimulating some part of their visual pathway, resulting in the perception of so-called ‘phosphenes’: small points in the visual field. However, the fidelity and resolution of such stimulations are currently limited. Thus, methods to make the greatest use of this limited bandwidth are necessary. To this end, a number of phosphene encoders have been developed. Additionally, in order to computationally evaluate these phosphene encodings, methods that attempt to extract useful information (such as the original image) from the phosphene image are necessary. In this thesis, I developed three such phosphene decoders as well as a di↵erentiable Canny edge detection encoder. I performed three experiments: one to compare the informativity of Canny edge detection encoding with object contour encoding, one to evaluate the e↵ect of phosphene resolution on informativity, and one to evaluate the e↵ect of Canny edge detection parameters on informativity, as well as the effectiveness of di↵erentiable Canny edge detection. I found that Canny edge detection yield better performance in decoding the original image, whereas object contour encoding yields better performance in terms of optical flow and semantic embedding decoding. Additionally, I found that phosphene resolution strongly and positively a↵ects decoding performance, except for semantic embedding. Finally, I found that clear patterns in the e↵ects of Canny edge detection parameters on decoding performance could be observed and that di↵erentiable Canny edge detection easily overfits the training data.
Faculteit der Sociale Wetenschappen