Corruption-Aware Multi-Scale Representation Learning for Robust Image Classification

Abstract

Deep neural classifiers often obtain high in-distribution accuracy while remaining sensitive to small distributional changes such as pixel noise, missing strokes, and local occlusion. This paper presents MSA-Net, a lightweight multi-scale attention network for robust image classification under constrained training budgets. Instead of processing an image only at its native resolution, MSA-Net builds three parallel representations from the original image and two average-pooled views. Each view is projected into a compact nonlinear embedding, and a learned scale gate computes an instance-dependent gated fusion before classification. The model is intentionally small so that the contribution of multi-scale fusion can be examined without relying on very large backbones. We evaluate MSA-Net on a reproducible MNIST protocol using 12,000 training examples, 2,000 validation examples, and 5,000 held-out test examples. Compared with compact single-scale and multi-scale baselines, MSA-Net trades a small amount of clean-set accuracy for substantially stronger robustness, obtaining 91.62% clean accuracy while maintaining 88.74% accuracy under Gaussian corruption and 85.78% under block occlusion. The results indicate that explicit scale gating is a simple but effective mechanism for preserving discriminative stroke information when the observed image is partially degraded.