A class shared dictionary learning method for image classification, has been published in Neurocomputing, 2019.
Created by Shuai Shao, Rui Xu, Weifeng Liu, Bao-Di Liu, Yan-Jiang Wang from China University of Petroleum.
We propose a novel dictionary learning algorithm named label embedded dictionary learning (LEDL). This method introduces the L1-norm regularization term to replace the L0-norm regularization of LC-KSVD. Compared with L0-norm, the sparsity constraint factor of L1-norm is unfixed so that the basis vectors can be selected freely for linear fitting. Thus, our proposed LEDL method can get smaller errors than LC-KSVD. In addition, L1-norm sparse representation is widely used in many fields so that our proposed LEDL method can be extended and applied easily. You can also check out paper for a deeper introduction.
In this repository, we release the code and data for training and testing.
In our paper, the proposed LEDL is compared with some traditional machine learning baselines
if you find our work useful in your research, please consider citing:
@article{shao2019label,
title={Label Embedded Dictionary Learning for Image Classification},
author={Shao, Shuai and Xu, Rui and Liu, Weifeng and Liu, Bao-Di and Wang, Yan-Jiang},
journal={Neurocomputing},
doi={10.1016/j.neucom.2019.12.071}
year={2019}
}
Firstly, you should download the datasets
Then, modify the number of training set and testing set.
options.tr_num = 10; % randomly select 10 samples per class as the training data
options.val_num = 5; % randomly select 5 samples per class as the training dataOur code is released under MIT License (see LICENSE file for details).