Quaternion Approximation Networks for Enhanced Image Classification and Oriented Object Detection

## Контекст Modern image classification and object detection tasks face significant challenges due to the need for rotation equivariance and efficient computation. Existing convolutional neural networks (CNNs) often struggle with maintaining geometric invariance to rotations, while traditional quaternion neural networks (QNNs) are computationally intensive and lack practical implementation. This paper addresses these issues by introducing Quaternion Approximation Networks (QUAN), a novel framework that combines the benefits of quaternion algebra with real-valued operations, ensuring efficient and rotation-equivariant processing of image data. ## Метод QUAN leverages quaternion algebra by approximating quaternion convolutions using Hamilton product decomposition. Instead of operating entirely in the quaternion domain, the network uses real-valued matrices to represent quaternion components. This approach ensures rotation equivariance while reducing computational overhead. Independent Quaternion Batch Normalization (IQBN) is introduced to stabilize training by addressing the unique challenges of quaternion-based layers. Additionally, spatial attention mechanisms are extended to quaternion operations, enhancing the model's ability to focus on relevant features. The framework is implemented with custom CUDA kernels to achieve high performance on modern hardware. ## Результаты QUAN is evaluated on standard benchmarks, including CIFAR-10, CIFAR-100, ImageNet for classification, and COCO and DOTA for object detection. Compared to traditional CNNs and other quaternion-based models, QUAN demonstrates superior accuracy with fewer parameters and faster convergence. For object detection, it achieves state-of-the-art (SOTA) performance among quaternion CNNs, showcasing its ability to handle rotation-sensitive tasks efficiently. The model's performance is attributed to its ability to preserve geometric properties while maintaining computational efficiency. ## Значимость QUAN holds significant potential across multiple domains. In robotics, its rotation-aware perception capabilities make it ideal for tasks such as autonomous navigation and object recognition. In other fields, its efficient architecture and ability to handle complex geometric transformations provide a competitive edge over conventional models. The framework's modular design and custom CUDA kernels ensure scalability and applicability to a wide range of real-world problems, including those requiring resource-constrained systems. ## Выводы QUAN advances the state-of-the-art in quaternion neural networks by introducing a novel approximation approach that combines the benefits of quaternion algebra with real-valued operations. Its superior performance in image classification and object detection, coupled with its efficient implementation, positions it as a promising solution for rotation-equivariant tasks. Future work will focus on extending QUAN to multi-modal data fusion and integrating it into larger modular frameworks for broader real-world applications.