SELECTING OPTIMAL PARAMETERS OF SLIC SUPERPIXELS BY USING DISCREPANCY MEASURES

Hits: 87
Research areas:
Year:
2017
Type of Publication:
Article
Keywords:
SLIC, Segmentation, Superpixel, OBIA, Discrepancy Measure
Authors:
Taskin Kavzoglu, Hasan Tonbul
Abstract:
Superpixels are homogeneous image regions consisting of spatially associated pixels. They produce meaningful image objects (segments) and provide less computational time for image processing tasks. Recently, the use of superpixels has become a popular research agenda in computer vision studies. Nevertheless, the operation of superpixel segmentation appears to be at a limited level in object based image analysis (OBIA). Segmentation that aims to partition an image into homogenous regions is the firstand most critical step of OBIA. There are many segmentation algorithms that utilize various parameters to control characteristics of output segments. However, selecting optimal parameter combination is a long-term and tedious process. In many studies, optimum parameter values are determined by discrepancy between a reference polygon and a corresponding segment as a segmentation evaluation criteria of image. In the ideal case, the expected situation is that over-segmentation and under-segmentation should be at a minimum level to achieve high-quality image segmentation. In this study, the effectiveness of Simple Linear Iterative Clustering (SLIC) superpixel segmentation algorithm was evaluated using varied parameter values for the generation of consistent image objects for land cover classes. Based on the discrepancy between reference polygons and corresponding image segments, theideal combination of SLIC parameter values were determined. In this context, five segmentation discrepancy metrics namely under-segmentation, over-segmentation, potential segmentation error (PSE), number-of-segments ratio (NSR) and Euclidean distance 2 (ED2) were applied through the manually digitized reference polygons to evaluate the segmentation quality of SLIC superpixels. A Worldview-2 and Quickbird-2 images of Turkey were used in the case study and four superpixel sizes (5x5, 10x10, 15x15, 20x20) were evaluated. Results show that superpixel sizes of 10x10 pixels produced the highest accuracy to identify optimal combinations of parameter values.
Full text: 191.pdf [Bibtex]
You are here: Home ACRS ACRS Overview Proceedings selecting-optimal-parameters-of-slic-superpixels-by-using-discrepancy-measures