Multi-Scale Unrectified Push-Pull with Channel Attention for Enhanced Corruption Robustness

Convolutional Neural Networks (CNNs) have achieved remarkable success in computer vision tasks, however, they often experience substantial performance degradation when confronted with real-world corruptions such as noise, compression artifacts, and lighting variations. The original push–pull CNN (PP-CNN) architecture addresses this challenge by employing a biologically inspired mechanism that contrasts local excitatory (push) and broader inhibitory (pull) responses to suppress noise. In this work, we enhance the robustness of PP-CNN through three key modifications: (1) removing the half-wave rectification constraint to enable more expressive interactions between push and pull signals, allowing for richer linear feature enhancement; (2) introducing a dynamic channel attention mechanism that adaptively recalibrates feature responses by amplifying discriminative signals and suppressing noise-dominated channels; and (3) designing a multi-scale push–pull (MSPP) framework that searches for pattern consistency across multiple spatial resolutions, reinforcing the model’s ability to generalize under corruptions at varying scales. Our proposed enhancements introduce a stronger inductive bias toward learning scale-consistent features—a fundamental property of natural images that remains stable even under corruption—without requiring corruption-specific data augmentation. Comprehensive evaluations on the CIFAR-10-C benchmark demonstrate that the enhanced PP-CNN achieves improvements in robustness across diverse corruption types while maintaining competitive accuracy on clean data. Notably, the multi-scale variant delivers the best trade-off between robustness and clean data performance, demonstrating the effectiveness of exploiting multi-scale feature consistency for generalization to unseen common image corruptions.