Feature Compression Based on Discrete Cosine Transform for Handwritten Digit Recognition

Ali Mohammed Almohammedi, Irfan Ahmad

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

In this paper, a new feature selection method is developed and investigated on handwritten digits classification task based on Discrete Cosine Transform (DCT). In the preprocessing stage, raw images are passed through normalization, slant filter, and edge-detection filter as well as the proposed DCT algorithm to reduce feature which can lead to sparse and orthonormal data matrix. The CNN classifier can inherently do feature extraction in additional to handwritten digit recognition. The proposed techniques of slant filter and DCT transformation along with the CNN classifier achieved the best global accuracy of 99.54%, 99.12%, and 98.52% for 784, 400, and 144 features, respectively on the public MNIST dataset of handwritten digits.

Original languageEnglish
Title of host publicationProceedings - 2020 1st International Conference of Smart Systems and Emerging Technologies, SMART-TECH 2020
EditorsAnis Koubaa, Ahmad Taher Azar, Basit Qureshi
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages13-20
Number of pages8
ISBN (Electronic)9781728174075
DOIs
StatePublished - Nov 2020

Publication series

NameProceedings - 2020 1st International Conference of Smart Systems and Emerging Technologies, SMART-TECH 2020

Bibliographical note

Publisher Copyright:
© 2020 IEEE.

Keywords

  • CNN
  • Discrete Cosine Transform (DCT)
  • Feature reduction
  • deep learning
  • slant and edge detection filters

ASJC Scopus subject areas

  • Electrical and Electronic Engineering
  • Instrumentation
  • Computer Networks and Communications
  • Artificial Intelligence
  • Energy Engineering and Power Technology
  • Control and Optimization

Fingerprint

Dive into the research topics of 'Feature Compression Based on Discrete Cosine Transform for Handwritten Digit Recognition'. Together they form a unique fingerprint.

Cite this