Date of Award
5-2024
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Computer Engineering
Committee Chair/Advisor
Melissa C. Smith
Committee Member
Jon C. Calhoun
Committee Member
Tao Wei
Abstract
JPEG (Joint Photographic Experts Group) was formed in 1986 to create a method to reduce image size primarily for ease of transfer on the Internet. Released to the public in 1992, JPEG compression is a form of lossless compression that has been a staple for compressing images. JPEG is the go-to image compressor because it provides high compression ratios while maintaining visual integrity for the human eye. Growing image sizes have made JPEG compression increasingly relevant. It is vital to keep up with growing data sizes for improved image handling performance on an edge device like a Field-Programmable Gate Array (FPGA). The JPEG compression format has had extensive research, but many potential applications have not yet been realized. Convolutional Neural Networks (CNNs) have become increasingly researched. CNNs boast the ability to interpret data across a variety of fields. In many cases, the data being collected is generated on a different device than the CNN. High-performance computers (HPCs) allow CNNs to run faster than on an edge device where data is collected. This is desirable, especially for more compute-hungry models that benefit from the additional performance provided by an HPC. This research aims to investigate improving the transfer of images to a more capable computer and the effect on accuracy on a few different Neural Networks due to a reduction in image quality. Compressing images to JPEG frees up bandwidth, allowing images to be transmitted faster since there is less data to transfer from memory. One of the primary investigations is to determine if JPEG images can be correctly interpreted by CNNs by comparing JPEG quality (various levels of compression and information loss) to CNN accuracy. Exploring the impact JPEG has on the runtime of the network will be analyzed to determine if the idea is viable. If successful, providing a universal method to improve the performance of a range of neural networks and the post-hoc transfer of the images while minimizing the reduction in model accuracy. Secondly, this research investigates developing the JPEG algorithm on an FPGA to determine the required resources of such an implementation and the performance benefits compared to the CPU variation. This investigation will demonstrate the ability to improve neural networks designed to run off data collected on an edge device.
Recommended Citation
Shaughnessy, Timothy, "Analysis of CNN Performance utilizing JPEG Compressed Images Created on an FPGA" (2024). All Theses. 4240.
https://tigerprints.clemson.edu/all_theses/4240