Publication Date

Fall 12-11-2019

Degree Type

Master's Project

Degree Name

Master of Science (MS)

Department

Computer Science

First Advisor

Robert Chun

Second Advisor

Melody Moh

Third Advisor

Chris Pollett

Abstract

The field of image recognition software has grown immensely in recent years with the emergence of new deep learning techniques. Deep belief networks inspired by Hinton [11] were one of the earliest methodologies of deep learning in the late 2000s. More recently, convolutional neural networks have been used in deep learning techniques, architecture, and software to identify patterns in imagery in order to make predictions such as classification, image segmentation, etc. Traditional two-dimensional, or 2D, images stored as picture files, typically contain red, green, and blue color data for each individual pixel in the picture. However, more recent commercial 2.5D or depth cameras have become more readily available such as the Microsoft Kinect, which is capable of capturing both RGB and depth (RGB-D) data. With the new depth dimension that can be captured from these cameras, objects are no longer limited to a flat dimension and the volumetric shape of the object can now be used to aid in recognizing that particular object.

In this project, I will utilize a convolutional deep belief network in order to observe the effects of rotation and sliding window stride when conducting classification on 3D models. An early study conducted named 3D ShapeNets experimented with this idea utilizing 3D computer aided design (CAD) model data in order to classify 3D models [2]. Extending from this research, the results from my research experiment showed an adverse correlation between angle granularity and recognition accuracy. Moreover, in regards to sliding window stride length, the training time increased substantially but had little effect on overall 3D model classification.

Share

COinS