Contextual Refinement of Neural Network Classification Using Relaxation Labelling Algorithms

Abstract. Neural networks are growing in popularity today as a tool for classification of remotely sensed images. Majority of the neural network applications involve the error backpropagation algorithm for training the network to work as a classifier. While this algorithm has been successfully employed in various image classification problems, the errors associated with per-pixel classifier still persist, even if on a smaller scale than other methods. The contextual information embedded in a pixel’s neighbourhood is a powerful mechanism to exploit the local knowledge and correct the errors. In this paper, the neural network output is scaled and input to the contextual classifier based on probabilistic relaxation labeling algorithm. The approach is tested using an IRS image and the results indicate that there is an improvement in classification accuracy over the conventional maximum likelihood method.