Edge-enhanced segmentation for SAR images

• Main Author: Ju, Chen
• Format: Thesis
• Language: English
• Published: Memorial University of Newfoundland 1997
• Subjects: Sea ice
• Online Access: https://research.library.mun.ca/1316/; https://research.library.mun.ca/1316/1/Ju_Chen.pdf; https://research.library.mun.ca/1316/3/Ju_Chen.pdf

Segmentation of Synthetic Aperture Radar (SAR) images is an important step for further image analysis in many applications. However, the segmentation of this kind of image is made difficult by the presence of speckle noise, which is multiplicative rather than additive. Traditional segmentation methods originally designed for either noise-free or White Gaussian noise corrupted images can fail when applied to SAR images. -- Different methods have been previously developed for segmenting SAR images corrupted by speckle. One segmentation method was proposed by Lee and Jurkevich which is quite efficient; it first smooths speckle noise to allow regions to be distinguished in the image histogram, then uses histogram thresholding to segment the filtered image. However, some problems exist with their method: in the filtered image, noise is preserved in edge areas and some fine regions are oversmoothed; while in the segmented image, region boundaries are ragged and some fine features are lost. -- Based on Lee and Jurkevich's initial work, an edge-enhanced segmentation method is proposed in this thesis. The edge-enhanced segmentation method is automated and based on the iterative application of an edge-enhanced speckle smoothing filter. The edge-enhanced filters proposed in this thesis use edge information obtained by a ratio-based edge detector to improve the performance of the filters in noise smoothing as well as in edge and fine feature preservation. Due to the good performance of these edge-enhanced filters, the resulting histogram-thresholded segmented images have accurate and simple region boundaries and well separated regions of both large and small sizes. The proposed method is compared with the previous method proposed by Lee and Jurkevich, in both noise smoothing performance and in segmentation quality. The results are tested on synthetic images as well as airborne SAR images. The tests show that the proposed method produces better image segmentations, particularly in image region boundaries, homogeneous regions and for images with fine features. The proposed edge-enhanced segmentation scheme may be suitable for many SAR image analysis applications such as sea-ice segmentation, forest classification, crop identification, etc.