.png){kind=icon}
.png){kind=icon}
Return
Regularized autoencoder based discriminative least square regression for image classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2025.113380.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Because of its efficient performance, least squares regression (LSR), which utilizes a discriminative projection for image classification, has gained significant attention. However, these two deficiencies limit its performance in real-world applications. Initially, existing LSR-based methods only consider global or local data, thereby reducing the robustness of the model. Secondly, the similarities within the same class are often neglected. Thus, we propose a regularized autoencoder based discriminative least squares regression (RADLSR) method to improve image classification. First, we integrate label relaxation LSR with an encoder-decoder framework to preserve global information in projection learning. Second, we consider the local relationship of the data as a complement to the relaxed autoencoder, helping to explore the complete structural information. Finally, we directly constrain the relaxed labels to be similar within the same class and incorporate the epsilon-dragging technique. This approach enhances the intra-class compactness and inter-class separability of the projected samples, resulting in a robust and highly discriminative model In addition, we develop an iteration optimization algorithm based on the alternating direction method of multipliers (ADMM) to solve RADLSR. Experiments on various benchmark datasets demonstrated that our RADLSR method can achieve superior classification results compared to state-of-the-art methods.
Keywords:
Regularised autoencoder
Discriminative regression
Intra-class compactness
Image classification