Exploring Edge Detection Methods for Improved Image Classification

Spitaels, Juliette

Etd

Exploring Edge Detection Methods for Improved Image Classification

Public Deposited

This work explores the best practices for edge detection in the context of a deep learning method of quantitative analysis, that is, using neural network accuracy as an objective measure of edge detection performance. To do this we first coded and implemented a variety of popular edge detection methods with an average-based automatic threshold selection algorithm. Then we specifically explored whether blurring an image to reduce noise and false edges is a beneficial step in edge detection in the context of image classification using neural networks. Additionally we tested if applying histogram normalization was a better alternative to blurring in the context of machine learning. Finally, we investigated the different effects of type I and type II errors in edge detection methods in the context of deep learning. The results indicated that in the context of simpler classification problems blurring did not yield significant improvement to neural network accuracy, but histogram equalization did improve network accuracy. In the context of more difficult classification problems, with more categories and fewer images per category, these results were no longer supported by the data. We also concluded that false-positive edges are preferable to false-negative edges when preprocessing images for classification using deep learning. Finally, we concluded that neural networks could be a valid method for quantitatively assessing edge detection methods, but with the drawback that randomization in the learning process does not guarantee a standard result, and this method therefore requires repetition to manage the variation in results.

Creator