Manuscript received August 27, 2024; revised September 26, 2024; accepted November 16, 2024; published April 25, 2025.

Abstract—Neural networks have been widely used in image recognition tasks, and this study explores a novel method— Quaternion Neural Networks (QNNs)—for enhancing performance. Using quaternion algebra, QNNs minimize the number of trainable parameters, leading to more compact models and quicker training times than Convolutional Neural Network CNNs. Therefore, color layers might potentially improve network performance by learning common parameters through input as linked values. Experiments assess learning processes by taking into account the roles of color and structure as well as stability in the presence of noisy visuals. According to the experimental results, QNNs retain an accuracy of 85% in the absence of noise, but at a noise level of σ = 0.30, accuracy dropped to 70%. Notwithstanding this, the network proved to be effective in learning structural information, exhibiting robustness against noise and disturbances in texture and color, hence confirming its suitability for wider image recognition uses. The paper establishes a proof of concept for the effectiveness of quaternion networks that will open up new avenues for research and possible uses that could outperform or supplement traditional networks.

Keywords—convolutional neural network, color, image recognition, quaternion neural networks

Cite: Taha Y. Abdulqader, Shahla A. Abdulqader, and Shatha A. Baker, "Enhancing Image Recognition with Quaternion Neural Networks: A Novel Approach to Color Layer Integration," Journal of Image and Graphics, Vol. 13, No. 2, pp. 190-197, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.