COLOUR IMAGE PROCESSING USING MODIFIED QUATERNION NEURAL NETWORK

Abstract

Convolutional neural networks are becoming more popular as a means of resolving issues related to the extraction of colour picture features. The inherent advantage of quaternion neural network had led to a recent upsurge in research related to its implementation in image, voice and signal processing. Incorporating quaternion algebra to a neural network can decrease the neural parameters and in spite of decreased parameters, it still achieves state-of-the-art performance. On the other hand, the interconnection of the colour image channels is not taken into consideration in the general network. Because of this, the authors of this research suggest a newly designed quaternion convolutional neural network (QCNN), which always handles colour triples as a whole in order to prevent the loss of information. In order to completely combine the data from the different colour channels, the first quaternion convolution process has been created and shown. In order to provide an even higher level of protection for the accuracy of colour information, the quaternion batch normalisation and pooling processes are developed and designed in the quaternion domain. During this time, information on the attention mechanism is being included into the proposed QCNN in order to improve its overall performance. Experiments show that the proposed model is superior in terms of efficiency to both the conventional convolutional neural network and another QCNN with the same structure. Furthermore, the proposed model demonstrates superior performance in terms of colour image classification and colour image forensics.