Color Image Recognition Models based on Quaternion Residual Networks

Color image recognition based on real-number neural networks can encounter challenges such as redundant self-information and insufficient mutual information in ternary color feature extraction. These issues can lead to inaccurate perception of color information by the model. Furthermore, deep neural networks with over 100 layers may experience a decline in model performance due to the large number of parameters. To overcome these challenges, this study extends real-number neural networks into the quaternion domain to construct a series of color image recognition models based on quaternion residual networks. The model designs include a lightweight structure, quaternion residual blocks, quaternion bilateral convolution, and quaternion L1 regularization. The quaternion residual network has half the number of parameters as the corresponding real-number network, and the lightweight design for parameters helps solve the computational complexity problem of deep networks. Moreover, the quaternion bilateral convolution can reasonably preserve the image color structure, and the quaternion L1 regularization can effectively sparse the training parameters, thereby mitigating network overfitting. Experimental results confirmed that the proposed models can improve image recognition rates while reducing algorithmic complexity compared to the corresponding real-number networks.