Review — Comparative Study of Classifiers for Blurred Images (Blur Classification)

Feature Extraction Using DCT, Classification Using RF Outperforms Naïve Bayes, MLP, k-NN, SVM

In this story, Comparative Study of Classifiers for Blurred Images, (Gueraichi SAI’20), by Houari Boumediene University of Science and Technology, is reviewed. In this paper:

  • A robust learning model based on the transformation into DCT is built, for the classification of images according to their degree of blur.
  • Naïve Bayes, MLP, k-NN, SVM, RF are tried. Finally, Random Forest (RF) is found to be the best.

This is a paper in 2020 SAI. (Sik-Ho Tsang @ Medium)

Outline

  1. Features Extraction Using DCT
  2. Experimental Results

1. Features Extraction Using DCT

Representation of (8 × 8) DCT bloc Frequency bands
  • An image of the Gblur dataset (from Live database) is divided into blocks of size (8×8) and on which the discrete cosine transform (DCT) is applied.
  • Each block obtained is divided (in addition to the DC band) into three regions R1, R2, and R3 which represent the low frequency (LF), medium frequency (MF) and high frequency (HF) regions, respectively.
  • On each frequency range, local statistics (mean, variance, kurtosis, skewness, entropy and energy, of all AC components constituting each region) are calculated.
  • The vector descriptor of the image consists of the averages of respective statistics found in all the blocks of the image.
  • The descriptor vector will therefore be composed of 18 features.
  • (6 features/region × 3 Regions = 18 features)
  • A test image (I) is classified as “slightly blurred”, “moderately blurred” or “strongly blurred” according to the three labels based on the Differential Mean Opinion Score (DMOS).
  • I is slightly blurred if 19 < DMOS ≤ 30.
  • I is moderately blurred if 30 < DMOS ≤ 60.
  • I is strongly blurred if DMOS > 60.

2. Experimental Results

  • Naïve Bayes, MLP, k-NN, SVM, RF are tried.
  • k-fold cross validation with k = 10, is used.
  • Within k-NN, it is found to be the best when k=3.
  • The chosen database of blurred images is Live’s Gblur database, which consists of 145 blurred images. Applying the class selection criteria using DMOS given by above, 39 low blur images (class 1), 78 medium blur images (class 2) and 28 high blur images (class 3).
Cohen’s kappa-statistics and G-score (Geometric mean) values for the five classifiers

RF is found to be the best. MLP is the second. SVM is the third.

  • But the training details has not been shown in the paper.

DCT+RF outperforms a SOTA approach NIQE.

  • (NIQE is a famous no-reference quality assessment approach.)

--

--

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store