Decoding Invisible 3D Printed Tags with Convolutional Neural Networks

Abstract

Imperceptible tags embedded on three-dimensional (3D) objects have recently shown promising utility in applications such as augmented and virtual reality interactions, tracking logistics, and robotics. The InfraredTag is a newly developed tag that is imperceptible to the eye and can be 3D-printed as part of an object. The InfraredTag can be detected by an infrared (IR) camera. A common problem with IR images is insufficient resolution, which may render the embedded tag unreadable, and image processing is required to increase contrast. Current image processing techniques use a different set of parameters for each filter and can take several seconds to finish, making it challenging to read InfraredTags in real time. To reduce processing time, the proposed thesis seeks to eliminate the need to try out all sets of parameters. It will instead use convolution neural networks (CNNs) to quickly convert an IR image into a binary image, from which the embedded code can be readily read.