Retracted Chapter: Non-Peaked Discriminant Analysis For Image Representation

The L1-norm has been used as the distance metric in robust discriminant analysis. However, it is not sufficiently robust and thus we propose the use of cutting L1-norm. Since this norm is helpful for eliminating outliers in learning models, the proposed non-peaked discriminant analysis is better able to perform feature extraction tasks for image classification. We also show that cutting L1norm can be equivalently computed using the difference of two special convex functions and present an efficient iterative algorithm for the optimization of proposed objective. The theoretical insights and effectiveness of the proposed algorithm are verified by experimental results on images from three datasets.