When a healthy human sees an object, the brain is capable of combining a right and a left visual version of the world. This process involves predicting what the left eye sees based on what the right eye sees and vice versa. The present dissertation will examine such processes using an autoencoder, a convolutional neural network with an encoding and decoding process.
This topic is approached by modelling qualities of the human retina including photoreceptor cells. Biological qualities of cones, the photoreceptor cells responsible for colour vision, are considered in order to adapt these processes. The performance of the model is assessed by looking at the case of double-vision, also called diplopia. The main objective is to understand which features of the retina help or do not have an effect on the correction of diplopia. Ultimately, the aim is to identify retinal features that support the capability to see a single version of an object.
A series of experiments is conducted by training the model on a dataset modified to account for the human retina. It will be demonstrated that it is in principle possible to correct diplopia using an autoencoder, but further extensions will be necessary to account for the effects of distribution and amount of cones.