Publication Date

12-2016

Date of Final Oral Examination (Defense)

10-27-2016

Type of Culminating Activity

Thesis

Degree Title

Master of Science in Computer Science

Department

Computer Science

Supervisory Committee Chair

Tim Andersen, Ph.D.

Supervisory Committee Member

Edoardo Serra, Ph.D.

Supervisory Committee Member

Aaron Westcott

Abstract

This work investigates the feasibility of using non-destructive testing, in particular radiation tomography, to recover length statistics from aggregates of fried batonnet cut potato. Non-destructive testing comprises a variety of useful techniques for determining properties of an object that might otherwise require altering or destroying the object physically. Tomography is a common form of non-destructive testing used primarily to infer properties internal to an object. This process involves exposing the object of interest to radiation and detecting the quantity of radiation energy that penetrates the object, usually resulting in a grey scale image.

To do this, an artificial data pipeline is developed in order to obtain annotated examples. This pipeline allows faster data collection than can be done in a real production environment coupled with the ability to control all aspects of the resulting images. Additionally, these examples are used to train a convolutional neural network, a widely successful machine learning algorithm for image processing. The network learns a relation between the images and the length estimates and can then be used to provide length estimates on novel examples.

Results show that with careful preparation and enough expected variation in the product being inspected, the image resulting from radiation tomography contains enough information to recover estimates of the lengths of the product with significantly less expected error than a naive baseline.

bsu-ms.cls (59 kB)
documentclass file for the tex thesis document

Share

COinS