Title: Attaining Rotation-Invariant Convolution Through Non-Learnable Orientation Alignment Operators

Abstract: Ensuring rotational invariance in deep neural networks without relying on data augmentation has emerged as a significant area of interest. Intrinsic invariance allows features to encapsulate the fundamental characteristics of targets, thereby boosting the efficacy of deep learning models in visual applications. Drawing upon a variety of non-learnable operators, this study introduces a robust framework of convolutional operations that are inherently invariant to any degree of rotation. Distinct from many existing approaches, these rotation-invariant convolutions (RIConvs) maintain an identical count of learnable parameters and a comparable computational workflow to conventional convolutions, allowing for seamless interchangeability. We demonstrate their rotational invariance across different angles using the MNIST-Rot dataset and benchmark them against earlier rotation-invariant CNNs, noting that two gradient-based RIConvs attain state-of-the-art performance. Furthermore, by embedding RIConvs into established CNN backbones, we assess their impact on texture recognition, aircraft classification, and remote sensing image categorization. The findings indicate that RIConvs substantially boost accuracy, especially when training data is scarce, and continue to enhance performance even in scenarios involving data augmentation.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC