| In image processing applications, finding a useful initial association among pixels to simplify higher level processing typically requires setting many parameters and making many ad hoc modeling decisions. Instead, the image could be segmented into distinct, homogeneous regions, whose use can improve performance in further processing, as well as reduce the amount of computation in later applications. To find these regions, an algorithm is developed, whose use only requires determining the scale at which these regions will be found. The algorithm adaptively determines the best predicates along which to group pixels for that scale. To make this possible, different class separability measures are developed and proven useful in the context of image processing. The results show that the output is comparable to that of the Mean Shift algorithm, without having to set any parameters. Thus, the ease of use of this algorithm along with its high quality results make this segmentation a suitable feature space for further processing. |
