Low-rank discriminative least squares regression for image classification

•A low rank discriminative least square s regression (LRDLSR) model is proposed.•LRDLSR aims at improving the intra class similarity of the regression labels learned by the” ϵ dragging technique involved in DLSR.•LRDLSR can ensure that the learned labels are not only relaxed but also discriminative.•LR DLSR also introduces a regularization term to avoid the overfitting problem by restricting the energy of learned labels.