Selecting optimal SLIC superpixels parameters by using discrepancy measures

Segmentation is the first and most crucial step of object-based image analysis in that an image is partitioned into homogenous areas, known as superpixels, considering the spectral, textural and contextual information of contiguous pixels. Superpixels have become popular for use particularly in computer vision. By considering superpixels instead of pixels redundancy and complexity in processing stage are minimized. In this study, Simple Linear Iterative Clustering (SLIC) superpixel segmentation method was evaluated for the generation of image objects by varying parameter values to search the optimal values. Based on the discrepancy between reference polygons and corresponding image segments, the ideal combination of SLIC parameter values were determined. To evaluate the segmentation quality of SLIC superpixels, five discrepancy metrics, namely under-segmentation, over-segmentation, potential segmentation error (PSE), number-of-segments ratio (NSR) and Euclidean distance 2 (ED2) were applied by considering manually digitized reference polygons. A Worldview-2 and a Quickbird-2 image covering two test sites from Turkey were employed, and four superpixel sizes (5x5, 10x10, 15x15 and 20x20) were evaluated to test the image objects quality. Results showed that the proposed metrics used to identify optimal combinations of parameters revealed the optimal size of superpixels size as 10x10 pixels. It was also observed that over-segmentation and the process time can be reduced by selecting the appropriate superpixel parameters.