.png){kind=icon}
.png){kind=icon}
Return
An interpretable spatial feature selection for dimensionality reduction in image classification
.png){kind=icon}
PRE
AI
DOI：10.1007/s10044-026-01654-7.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Feature selection is widely used to reduce data dimensionality by identifying and retaining the most crucial features. It is often used to save the inference time of neural networks. This study introduces a new feature selection framework to reduce dimensionality without compromising classification accuracy. It leverages a tailored model to enforce a linear relationship between the model's backbone outputs and classifier outputs, enabling class-wise one-versus-rest linear separation. The structural design enables the mathematical interpretation of how selected spatial feature regions contribute to the model's final decision, providing theoretical justification for the feature selection process. The proposed method uses support vector machines to identify class-separable subspaces in the backbone outputs. Among those subspaces, the smallest one is then projected back to the original input space, and the corresponding image region is preserved. The effectiveness of the method was validated on various image datasets. Experimental results demonstrate that the proposed method effectively reduces dimensionality by up to 67% with minimal degradation in classification accuracy, at less than 1.2%. Moreover, an ablation study suggests that the method can control the trade-off between dimensionality reduction rates and classification accuracy.
Keywords:
Feature selection
Dimensionality reduction
Image classification
Linear separation
Journal
IF:
2
Papers:
1.9K
Citations:
1.9K