Computer Information Systems and Engineering Technology


Conner Saeli



Download Full Text (2.4 MB)


Conner Saeli, CIS 494: Research in Computer Information Systems
Faculty Mentors: Professor Gang Hu, Computer Information Systems and Professor Sarbani Banerjee, Computer Information Systems

The goal for this research was to develop a novel edge detection method. This project originated from studying new edge detection methods using convolutional neural networks. Edge detection provides a global view of an image with the most critical outline. Robust edge detection methods would allow for concrete analysis in image processing. The applications for this are useful in a wide range of fields, from satellite imaging to medical screening. For example, medical imaging results could be examined in batches to detect anomalies without any human intervention. Modern deep neural network-based edge detection methods achieve roughly 80% accuracy, so the intended result was to improve on this baseline. This model is based on a PyTorch implementation of a current edge detection algorithm called Holistically-Nested Edge Detection (HED). HED's approach is to calculate the loss and update its weights using not only intermediate feature maps, but also with a single added result of each feature map called a fusion layer. Instead of adding intermediate feature maps together from the deep network, this approach adopts multiplication into the network. The objective is to extract more information from network layers than possible with existing approaches. By multiplying instead of adding, this project aims to find agreement between intermediate feature maps. This implementation of HED uses pre-trained parameters from the VGG-16 convolutional neural network as a starting point. The datasets used for this project are BSDS500 and NYUDv2, which are public datasets from UC-Berkeley and New York University, respectively.

Publication Date



Computer Sciences | Engineering

Multiplicative Agreement Edge Detection
Off-campus Download