Files in this item



application/pdfECE499-Sp2018-snyder.pdf (1MB)
(no description provided)PDF


Title:Classification centric Lloyd-Max quantizer for predictive feature errors
Author(s):Snyder, Corey
Contributor(s):Moulin, Pierre
Subject(s):Classification Centric Lloyd-Max Quantizer
image and video compression
compression ratio
Abstract:Conventional forms of image and video compression have a singular objective to optimize visual image quality in the presence of a target compression ratio. However, if we would like to perform machine analysis of these compressed images, what we as humans view as interpretable is distinct from what machines may find informative. This paper will discuss compression techniques with the dual purpose of maintaining image quality and preserving image features for machine classification. The format of our system is a two-part predictive encoder. Features are extracted from both the original and JPEG 2000 compressed images. The difference between these feature vectors are the resulting error vectors due to compression. We seek to build an effi cient quantizer to encode the predictive feature errors. We evaluate the effectiveness of our quantizer by examining its ability to recover accuracy lost in an image classifier due to compression. We present the Classification Centric Lloyd-Max Quantizer as an effi cient vector quantizer to restore feature vector integrity and maintain high compression ratios. We compare previous work that utilizes scalar Lloyd-Max Quantizers and gradient descent methods to build an optimal vector quantizer against our proposed system. We demonstrate the advantages and disadvantages of each approach with respect to classification accuracy, compression ratio, training time, and constraints on the classification problem.
Issue Date:2018-05
Date Available in IDEALS:2018-05-24

This item appears in the following Collection(s)

Item Statistics