Contextual image labelling with a neural network

Abstract

A neural network with a multilayer perceptron architecture is shown to be capable of labelling the visible objects in colour images of urban and rural outdoor scenes. The two problems of segmentation and recognition are separated by using 'ideal' segmentations, allowing the performance of the recognition method to be studied independently of the effects of using an imperfect real segmentation process. A label clustering transformation is proposed and shown to cause a significant increase in the expected classification accuracy of the network. The deletion of the contextual features from the feature vector is shown to degrade the performance of the network. Measurements of the generalisation performance on unseen test data show that, on average, the system correctly recognises approximately 72% of the area of these images.