Eliminating Global Positions within Convolutional Neural Networks

Presenter

Jacob Rottenberg

Campus

UMass Amherst

Sponsor

Weibo Gong, Department of Electrical and Computer Engineering, UMass Amherst

Schedule

Session 2, 11:30 AM - 12:15 PM [Schedule by Time][Poster Grid for Time/Location]

Location

Poster Board A9, Campus Center Auditorium, Row 1 (A1-A20) [Poster Location Map]

Abstract

Computer Vision has not seen the major success that other forms of AI and deep learning models have. This is in part due to the approach that such models employ where they utilize the entire image and global position information within the image. This deviates from nature and the founding ideology of the neural network. The purpose of this research project is to propose a new type of deep learning layer that is able to extract essential information about an image. This ability should allow the recognition and transformation of objects to be much easier. To achieve this, the image is broken down into patches so that global coordinate information is not used. Then using these small patches, their invariant information is processed such that when reconstructing the image and mapping it to the new image, the patches’ location can be tracked. I was able to make this framework in Python using different mainstream libraries like Tensorflow, Keras, Numpy and Scipy. The overall conclusion of this project is that processing images using relative coordinates rather than global ones is viable given enough computing resources and training data. The preliminary results show that invariant information is learned and is transferred via the constructed mapping matrix. These results represent a significant process change in the world of image and video processing. Utilizing the relative coordinate technique may allow for quick processing of image and future prediction within a live-feed video. This has implications for self-driving cars and active predictions in other industries.

Keywords

Artificial Intelligence, Convolutional Neural Network, Computer Vision, Deep Learning, Relative Image Coordinates

Research Area

Artificial Intelligence

SIMILAR ABSTRACTS (BY KEYWORD)