Optimized image segmentation and its effect on classification accuracy

Image segmentation is a preliminary and critical step in segment-based image analysis. Its proper evaluation ensures that the best segmentation result is used in image classification. In this paper, image segmentations were carried out and the results were evaluated with an objective function that aims at maximizing homogeneity within segments and separability between neighbouring segments. The segmented images were classified with Maximum Likelihood Classifier (MLC) and the classification results were evaluated with independent ground data. The optimal segmentation, i.e. with the highest objective function value, also resulted in the highest classification accuracy, which is 5.92% higher than that obtained by the segmentation with the lower objective function value, and the difference is significant by McNemar’s test with p= 0.05, p is the significance level. This shows that the objective function is indeed an effective way to determine the optimal segmentations to carry out the classifications. Pixel-based MLC was also carried out to compare with the segment-based classification. Besides free of salt-and-pepper effect, the best-segmentationbased classification obtained accuracy 2.3% higher than obtained by the pixel-based classification. Though by McNemar’s test, the difference is not significance, with p=0.05. This result seems to suggest that the benefit of segmentation-based classification lies not only in the segmentation step, which alone leads to marginal classification improvement, but that the use of segments’ shape, contextual as well as spectral information, is needed to increase accuracy significantly.