A Quantitative Evaluation of Image Segmentation Quality

Segmentation is often a procedure to group spatially adjacent image pixels into segments. Spatially adjacent pixels form one image segment if they meet some criteria, such as spectral similarity. A segmentation result may vary, depending on a given generalization level and other constraints such as compactness. In practice, the quality of a segmentation result is often assessed visually by the analyst, and that lacks a quantitative support and the quality of the result relies on the experience of an analyst. This research gives a quantitative estimate of a segmentation result using indices such as: a) a summed standard deviation of the input images within each image segment; b) a summed absolute difference within each image segment; and c) a summed difference of a segment to its adjacent segments. A series of segmentation results were studied from two perspectives. One focused on a series of segmentation results with different generalization levels, the other compared two series of segmentation results obtained using two different methodologies. The findings of this study can be used a) to guide an algorithm optimization for image segmentation, b) to assist the evaluation of different software programs to be used for a certain application and, c) to assist in identifying an optimal segmentation result for a given analysis.