Abstract
Neural networks in the real domain have been studied for a long time andachieved promising results in many vision tasks for recent years. However, theextensions of the neural network models in other number fields and theirpotential applications are not fully-investigated yet. Focusing on colorimages, which can be naturally represented as quaternion matrices, we propose aquaternion convolutional neural network (QCNN) model to obtain morerepresentative features. In particular, we redesign the basic modules likeconvolution layer and fully-connected layer in the quaternion domain, which canbe used to establish fully-quaternion convolutional neural networks. Moreover,these modules are compatible with almost all deep learning techniques and canbe plugged into traditional CNNs easily. We test our QCNN models in both colorimage classification and denoising tasks. Experimental results show that theyoutperform the real-valued CNNs with same structures.