CONTEXTUAL IMAGE LABELING WITH A NEURAL-NETWORK

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 recogn ition are separated by using 'ideal' segmentations, allowing the perfo rmance of the recognition method to be studied independently of the ef fects of using an imperfect real segmentation process. A label cluster ing transformation is proposed and shown to cause a significant increa se in the expected classification accuracy of the network. The deletio n of the contextual features from the feature vector is shown to degra de the performance of the network. Measurements of the generalisation performance on unseen test data show that, on average, the system corr ectly recognises approximately 72% of the area of these images.