Reducing the random seed effect on segmentation by applying an edge-preserving filter

Abstract

In region-growing segmentation algorithms random seed locations are used (reference). To ensure that repeating the segmentation will produce the same result, the seed locations are following a fixed random pattern. Empirical studies show that when the image that is subjected to the segmentation is changed by adding or removing rows or columns, the resulting segments are not identical anymore, the so-called seed effect. This occurs not only at the border of the image as one would intuitively expect, but also in the center part of the image. Apparently, the exact location of a seed affects the resulting segment. In this study I investigated whether application of an Edge-Preserving Smoothing Filter to an image prior to segmentation would reduce the effect of seed locations when rows or columns are added or removed from that image, and hence make the segmentation method by random seeds more robust. Two images were included: an IKONOS image of the central part of the Netherlands with fragmented land cover of agriculture, forest and villages and a SPOT5 2.5m multi-spectral image of a semi-desert steppe area in southeastern Kazakhstan. Both images were subjected to an Edge-Preserving Smoothing Filter before segmentation. This filter calculates variance for each band in nine different directions (8 wind directions + central area) and sums them per direction. The average band values of the direction with the lowest overall variance are then assigned to the central pixel. For both areas four subsets, each measuring 500x500 pixels, were selected representing different kinds of landscape with different patterns. For the IKONOS image the subsets covered a forested area with a golf course, a business area, a residential area, and an agricultural area. The subsets of the SPOT5 image covered a floodplain, a dune area with sparse vegetation where the soil is covered by lichens, a dune area with sparse vegetation but without lichens, and a dune area with many patches without vegetation and without lichens. In total 10x2 images were available for the analysis (8 subsets, 2 full images, original and EPSF). All images were segmented at five different heterogeneity levels. To quantify the effect of the seed location, the tessellations of the subsets were compared to the tessellation of the full image by overlaying and by analyzing the length of segment borders, both for the original and the EPSF versions. The length of segment borders that coincided between the subset and the full image was divided by the length of all borders in the subset (excluding the enveloping rectangle). The outcome was subtracted from 1 and this value was taken as a measure to quantify the seed effect. The results show that the segmentation results are more similar between the subset and the full image when the EPSF filter was applied before segmentation. In the heterogeneous area in the Netherlands, the seed effect was on average reduced by 6.6% when applying the EPSF filter. In the more homogenous area in the semi-desert in Kazakhstan, the seed effect was reduced by 48.8%. The seed effect strongly increases with higher heterogeneity levels, for all subsets and for both images. The explanation for the positive effect of the EPSF filter is likely found in the creation of very small homogenous patches. The seed pixels will be grouped with pixels from the same patch first, which reduces the location effect on the final segmentation.