A two-stage process for accurate image segmentation

Segmenting images is one of the most important steps in many high-level computer vision algorithms. The ability to divide images up into meaningful regions based upon properties such as shape, texture and colour has still not been fully solved. In this paper we show how good quality segmentations of complex outdoor scenes may be achieved by a two-stage process. The first step is to produce an approximate segmentation using only colour and texture information. The second step is to merge regions by using a neural network trained to classify the regions into one of eleven possible types which correspond to objects types found in outdoor scenes. This stage involves the use of high-level knowledge such as position, shape, context and orientation.