A Biologically Motivated Software Retina for Robotic Vision Applications

We present work in progress to address current limitations in image analysis by Deep Convolutional Neural Networks. By applying structural constraints based on known properties of the human visual system we propose to facilitate learning simple scale and rotation transformations, which contribute to large computational demands for training and opaqueness of the learned structure. We propose to apply a version of the retino-cortical transform to reduce the dimensionality of the input image space by a factor of e×100, and map this spatially to transform rotations and scale changes into spatial shifts. By reducing the input image size accordingly, and therefore learning requirements, we aim to develop a compact and lightweight robot vision sensor using a smartphone as the target platform. We also consider the visual processing architectural issues that must be addressed to integrate the mobile phone based front-end within a larger robot cognitive vision system.