Combining of Markov Random Field and Convolutional Neural Networks for Hyper/Multispectral Image Classification.

In the last decades, computer vision tasks such as face recognition, pattern classification, and object detection, have been extensively explored due to advances in computing technologies, learning algorithms, and availability of large amounts of labelled data. Image texture analysis is helpful and important to extracting valuable features of images in order to perform the aforementioned tasks. Markov Random Field (MRF) theory defines probabilistic relationships between the pixel value and its neighbouring pixels, which can help to produce features for classification purposes with deep learning. Convolutional Neural Network (CNN) is a popular feed-forward deep neural network, especially for handling visual tasks. Therefore, this study focuses on combining MRF and CNN in order to achieve hyper/multispectral image classification tasks. MRF images were generated to produce prefixed MRF filters for the first and/or second attention-like layers of CNNs to better extract features. Then CNNs were trained with the prefixed filters for experimentation on the UoM, BONN, and Cassava datasets. Experimental results have demonstrated that such combination of MRF and CNN can enhance classification accuracy while being more time-efficient for hyper/multispectral image classification compared to CNN structures without prefixed MRF filters.