Image classification is a central problem in computer vision and is most often solved using deep learning models. While these models achieve strong performance, they are typically large, complicated, and difficult to interpret. To address these limitations, we aim to explore an alternative paradigm: Green Learning, which focuses on building efficient and interpretable models.
One important direction of our work is introducing supervision into the feature extraction process. A key component of our approach is the LDA filter block, a feedforward mechanism that uses Linear Discriminant Analysis (LDA) to construct convolution filters without relying on backpropagation. These LDA filters align image patches with class-discriminative directions.
In addition, we propose spectral LNT, a new variant of the least-squares normal transform (LNT) that leverages the spatial–spectral structure of feature maps by applying localized linear combinations of features. The resulting LNT kernels can be naturally interpreted as LNT projections over local receptive fields. Moreover, we adopt a pyramid sparse-coding structure to extract sparse-coding features from a Gaussian pyramid of the input image. This further enriches the feature representation and leads to improved classification accuracy.
