realamirhe / smlfdl Goto Github PK

📄article

SVMs multi-class loss feedback based discriminative dictionary learning for image classification

SMLFDL integrates dictionary learning and support vector machines training into a unified learning framework by looping the designed multi-class loss term, which is inspired by the feedback mechanism in cybernetics.

analysis has been done on scene-15 dataset.
Feature vectors has been prepared by four-level spatial pyramid, dense DAISY feature description followed by PCA.
As article proposed SMLFDL are faster in predictions and converge in lower epochs.
_{code for features will be added soon.}

Highlights:

• Inspired by the feedback mechanism in cybernetics, a novel discriminative dictionary learning framework, named support vector machines (SVMs) multi-class loss feedback based discriminative dictionary learning (SMLFDL) is proposed to learn a dictionary while training SVMs. As far as we know, it is the first time that the feedback mechanism in cybernetics is adopted for constructing dictionary learning model.

• SMLFDL further employ the Fisher discrimination criterion on the coding coefficients under -norm constraint to make the coding coefficients have small intra-class scatter but big inter-class scatter for countering intra-class variability of datasets.

• An efficient and practical SMLFDL optimization algorithm is presented to learn a dictionary while training SVMs. Experimental results on several widely used image databases show that SMLFDL can achieve a competitive performance with other state-of-the-art methods on classification task.

Notes:

The original article was developed in matlab

The report file is an over-view showing precedures and some figures and didn't published anywhere, it must not be refernece any where, for refernece use original article