Ensemble SABA-Net: CPU-Efficient Lightweight Image Classifier for Resource-Constrained Environment

Deep learning has driven remarkable progress in visual recognition, yet state-of-the-art models remain heavily reliant on large-scale labeled datasets and high-performance GPU infrastructure, assumptions that rarely hold in real-world industrial settings where data is scarce and deployment occurs on CPU-based systems. This working paper introduces the Ensemble Spatial-Agnostic Basis Adaptation Network (Ensemble SABA-Net), a lightweight classification framework explicitly designed for low-data and resource-constrained environments. The proposed architecture departs from conventional convolutional and attention-based designs by operating directly on flattened pixel intensities, extracting hierarchical representations through shallow multi-layer perceptrons, and constructing class-specific decision boundaries via local prototype learning. Each ensemble member generates multi-layer embeddings that are clustered into prototypes per class, enabling distance-based classification. The ensemble mechanism aggregates predictions across independently initialized estimators to enhance robustness and reduce variance. Experimental evaluations on MNIST and Fashion-MNIST under severe data limitations (500 training samples) demonstrate that Ensemble SABA-Net achieves competitive accuracy while reducing training time by approximately 85% compared to Vision Transformers and maintaining inference times under 0.5 milliseconds per sample on CPU hardware. The framework converges faster, achieves higher final accuracy in low-data regimes, and eliminates dependence on GPU acceleration. These results establish Ensemble SABANet as a practical alternative for industrial applications such as defect detection and specialized visual analysis, where data poverty, computational constraints, and interpretability requirements necessitate alternatives to mainstream architectures. The work bridges the gap between cutting-edge vision research and the operational realities of resource-limited deployment environments.