Abstract

Deep learning has achieved great accuracy in large scale image classification and scene recognition tasks, especially after the Convolutional Neural Network (CNN) model was introduced. Although a CNN often demonstrates very good classification results, it is usually unclear how or why a classification result is achieved. The objective of this thesis is to explore several existing visualization approaches which offer intuitive visual results. The thesis focuses on three visualization approaches: (1) image masking which highlights the region of image with high influence on the classification, (2) Taylor decomposition back-propagation which generates a per pixel heat map that describes each pixel's effect on the classification, and (3) Inception which generates a natural looking image based on the features maximizing the classification score. We explore two challenging visualization tasks, (1) visualizing a model classifying images based on the time when they are taken, and (2) visualizing a model of predicting plant phenotypes (specifically wheat heading percentage). The thesis demonstrates how these visualization approaches work for both the classification model and regression model, and evaluates the results on real-world imagery.

Committee Chair

Robert Pless

Committee Members

Yasu Furukawa Tao Ju

Comments

Permanent URL: https://doi.org/10.7936/K7BK19NT

Degree

Master of Science (MS)

Author's Department

Computer Science & Engineering

Author's School

McKelvey School of Engineering

Document Type

Thesis

Date of Award

Spring 5-15-2016

Language

English (en)

Download

Share

COinS