Quaternion Convolutional Neural Networks for Image Classification

In recent years, real-valued neural networks have made significant progress in computer vision tasks such as image classification, object detection, and semantic segmentation. However, traditional convolutional neural networks (CNNs) struggle to effectively capture complex relationships between channels when handling multi-channel data with inherent correlations such as color images, medical images, and remote sensing data, limiting the model’s expressive power and generalization ability. To address this, we adopt the quaternion convolutional neural networks (QCNNs), which leverage quaternion algebra to create a more compact and efficient deep learning framework. QCNNs better capture the internal correlations in multi-channel data, reduces model parameters, lowers computational costs, and enhances feature extraction and model generalization. Experimental results show that QCNNs outperform traditional real-valued CNNs in color image classification, demonstrating higher efficiency and performance, with great potential for future research.