computational measure of texture similarity is not consistent with human perception
need perceptually consistent similarity measure
need to be invariant to intensity, contrast, scale, orientation
  
 

construct perceptual texture space (PTS)
map input texture feature to PTS through invariant space
  

  collect sample texture images
normalize texture intensity, constrast, scale, orientation
get human subjects to perform free sorting task
compute perceptual distance matrix from free sorting results
apply Multidimensional Scaling to obtain sample coordinates from distance matrix
   

MDS fitting result

(stress = fitting error)

Result: 4D is optimal

3D Perceptual Texture Space
 

 
VRML visualization of 3D PTS   (sorry, can't visualize 4D)
(require VRML plug-in, e.g., Cortona)
 

 

our perceptual texture space is consistent with existing spaces
       

computational texture features and distances are not consistent with PTS
 
         
  

  use convolutional neural network to perform invariant mapping
use SVM to perform perceptual mapping
 
         

input: Gabor features arranged in 2D
 

texture image
Gabor features in 2D arrangement
       orientation           
frequency       
      
  mapping performance
       Tamura, MRSAR, invariant features are mapped using SVM
       Ic = new texture instance; canonical scale, orientation
       Tc = new texture type; canonical scale, orientation
       Iv = new texture instance; variable scale, orientation
       Tv = new texture type; variable scale, orientation
       R = random

       NN+SVM produces best overall results

            

retrieval performance

       NN+SVM yields best retrieval performance

             
 
  A/Prof. Leow Wee Kheng, Dept. of Computer Science, National University of Singapore.
A/Prof. Chua Fook Kee, Dept. of Social Work & Psychology, National University of Singapore.

Ex-students:
Dr. Long Huizhong (Ph.D.), Dept. of Computer Science, National University of Singapore.
Mr. Tan Chee Wee (Honours), Dept. of Computer Science, National University of Singapore.
 
  H. Long and W. K. Leow. A hybrid model for invariant and perceptual texture mapping. In Proc. ICPR, 2002.
  H. Long and W. K. Leow. Perceptual consistency improves image retrieval performance. In Proc. ACM SIGIR, 2001.
  H. Long, C. W. Tan, and W. K. Leow. Invariant and perceptually consistent texture mapping for content-based image retrieval. In Proc. ICIP, 2001.
  H. Long and W. K. Leow. Perceptual texture space improves perceptual consistency of computational features. In Proc. IJCAI, 1391-1396, 2001.
  H. Long, W. K. Leow, and F. K. Chua. Perceptual texture space for content-based image retrieval. In Proc. MMM, 167-180, 2000.
     
  This project is part of the project supported by NSTB and MOE.