Fast labelling of natural scenes using enhanced knowledge (Q1606686)

A new technique for labelling natural scenes is proposed. This technique labels disjoint regions on an image of a natural scene on the basis of knowledge about the relationship among objects. The proposed technique consists of three stages: (1) segmentation, (2) initial labelling, and (3) label improvement. One of the most promising previous techniques uses simulated annealing to find the solution, while our technique uses local hill-climbing with enhanced knowledge for speeding up the processing. Local hill-climbing is known to be easy to be captured by a local minimum. We solved this problem by enhancing the knowledge being used as constraints for the search. Our knowledge represents 1-to-n relationships among regions, pair-wise relationships of regions, and relative locations of the regions to the image. In addition, we introduced two region features: an entropy in intensity; and a linearity of contours of each region. The linearity evaluation aims to distinguish artificial objects from natural objects. The validity of the technique is supported by some experiments. These experiments showed that the proposed technique is much faster with the almost same accurate.