textureclassification_BIGD_SPIC2020

Texture classification based on block intensity and gradient difference (BIGD) descriptor

Texture classification using block intensity and gradient difference (BIGD) descriptor (label: bigd_version1.0)

• Copyright: 2020 Yuting Hu
• Affiliation: Omni Lab for Intelligent Visual Engineering and Science (OLIVES), School of Electrical and Computer Engineering (ECE), - - Georgia Institute of Technology
• Email: huyuting.sjtu@gmail.com
• Citation: If you use the code and data please cite the following in your work:
• Y. Hu, Z. Wang, and G. AlRegib, “Texture classification using block intensity and gradient difference (BIGD),” Signal Processing: Image Communication, vol. 83, 2020.
• @article{hu2020bigd, title={Texture classification using block intensity and gradient difference (BIGD)}, author={Hu, Yuting and Wang, Zhen and AlRegib, Ghassan}, journal={Signal Processing: Image Communication}, volume={83}, year={2020}, publisher={Elsevier} }

Resources

Our code is built on top of following references/resources:

• dense microblock difference (DMD): the source codes of DMD was downloaded from http://www.cs.tut.fi/~mehta/texturedmd before. Paper reference: Mehta, Rakesh, and Karen Egiazarian. "Texture classification using dense micro-block difference (DMD)." In Asian Conference on Computer Vision, pp. 643-658. Springer, Cham, 2014. The implemnetation of block intensity and gradient difference (BIGD) part is based on the revision of DMD codes.
• VLFeat toolbox is used.
• K-means and vector of linearly aggregated descriptors (VLAD): the source codes of the proposed algorithm can be downloaded from W.

Objective, datasets, main functions, and key parameters

Objective:

Texture classification based on block intensity and gradient difference (BIGD) descriptor

Datasets:

KTH-TIPS, KTH-TIPS-2a, KTH-TIPS-2b, and CUReT under direcoty "data" Due to the big data size around 2GB, datasets are not included here but they can be automatically downloaded to direcoty "data" when "dmdVlad_demo.m".

Main functions

  Run bigdVlad_demo.m as the main function and we ran Matlab2014b codes on Intel Core i7-4790K 4.00GHz
  a. bigdVlad_demo: a demo for the texture classification using BIGD features
     - Initialization: dataset direcoty and parameter setting for BIGD and VLAD encoding stored in opts, dmdOpts, and opts, respectively
     - Outputs: mean average accuracy of texture classification over 10 cross validations   
  b. testDmdVlad(opts, DMDopts, FVopts):
     - Inputs: options for dataset and result diretories opts, options for BIGD feature extraction DMDopts, and optionss for VLAD encoding FVopts
     - Outputs: classification accuarcy on each cross validation
  c. trainEncoder(images, FVopts, DMDopts):
     - Inputs: texture images images, VLAD encoding options FVopts, and BIGD feature extraction options DMDopts
     - Outputs: a visual word dictionary using k-means denoted by centers
  d. computeIGradientDmd(img, dmdOpts)
     - Inputs: input image img, BIGD feature extraction options dmdOpts
     - Outpus: local BIGD descriptor denoted by features
  e. encodeVlad(centers, im, DMDopts, FVopts):
     - Inputs: a visual word dictionary centers, texture images im, BIGD feature extraction options DMDopts, and VLAD encoding options FVopts
     - Outputs: image descriptors descrs
  f. [W B] = vl_svmtrain(X, Y, LAMBDA): a linear SVM trained from the data vectors and the labels
     - Inputs: the data vectors X, the labels Y, and lambda parameter in the objective function LAMBDA
     - Outputs: a weight vector W and offset B
  g. [RECALL, PRECISION, info] = vl_pr(LABELS, SCORES): computes the precision-recall(PR) curve
     - Inputs: the ground truth labels LABELS and the scores of the samples obtained from a classifier SCORES
     - Outputs: info
     info.ap and nfo.ap_interp_11: average precision and 11-points interpolated average precision as defined by TREC

Key parameters

 - datasetList: the name of database
 - blkRadii: the radius of a local patch
 - nPoints: the number of sampling pairs in a local patch at each scale
 - gridspace: the spacing of sampled center pixels for each local patch
 - scale #: the number of scales
 - scale size: the size of block pairs with a local patch
 - numDescrs: the total number of randomly sampled local descriptors for each image
 - numKmeanscluster: the number of k-means clustering centers
 - crossValIndex: the number of cross validations
 - numTrain2 and numTest: the corresponding number of traning images and test images for each class
 - Normalization after encoding: flag marking if there is a normalization step after image encoding or not
 - meanAcc: mean classification accuracy
 - pmAcc: standard deviation of classification accuracy