A Quaternion Deep Learning Model for Color Image Classification

In recent years, residual network research in the real number domain has shown promising results in various computer vision tasks. However, the po-tential applications of residual network models in other numerical domains have yet to be fully explored. For color images, feature extraction in the real-number domain fails to preserve the structural information of ternary colors. The information loss will significantly impact the efficiency of residual network models. To solve the above problems, considering the enhanced multidimensional feature representation capability of quaternions in signal processing, we introduce a new quaternion domain feature-based residual network in this study. The quaternion deep learning model is constructed by designing the model structure, and quaternion convolution, pooling and regularization algorithms for the residual network. This construction resulted in an enhancement in image classification rates and the model's generalization capacity, coupled with a reduction in network overfitting. Comprehensive experiments were conducted on the Cifar-10, Cifar-100, and Oxford 102flowers datasets with test classification accuracy rates of 94.63%, 83.36%, and 97.26% respectively. The experimental results confirm that the proposed model is more efficient than the corresponding real-valued residual network in the task of color image classification. This conclusion provides an effective strategy for describing color information in images and extending residual networks to the hyper-complex number domain.