Learning from Imperfect Reconstruction: Data Augmentation via Dual Convolution

Data augmentation is a critical technique for improving the generalization and robustness of Convolutional Neural Networks, especially when training data is limited. Existing methods, however, often face challenges like high computational cost and semantic inconsistency. To address this, we propose Dual Convolutional Augmentation (DCA), a novel, lightweight, and efficient method that requires no additional training. DCA sequentially applies a convolution and a transposed convolution with a shared kernel to an input image. This process introduces structured, localized perturbations that simulate realistic degradations while preserving core semantic features. Comprehensive experiments on four public datasets demonstrate that DCA yields superior performance compared to traditional and several state-of-the-art augmentation techniques, particularly in enhancing model robustness against Gaussian noise and blur. Our analysis further reveals that features learned with DCA exhibit better clustering properties, confirming its effectiveness in model regularization.