Morphological Extraction of Line Networks from Noisy Low-Contrast Images

Abstract Novel algorithms for line and line network extraction from images are proposed. The algorithms which operate on region adjacency graphs are developed especially for images that are noisy and have low contrast. Microscopic images of fibers are used as an example application. The proposed algorithms are very general in nature, since they can detect lines of arbitrary geometry. The region adjacency graph is constructed from a segmented version of the original image. The segmentation process is based on the watershed transformation which is widely used in morphological image analysis. A Laplacian-type operation for line detection on the region adjacency graph is presented and removal of noise structures is studied. Fast algorithms for the detection of line ends and line branches are proposed. In addition, a directional propagation algorithm is derived. The line-end and branch-detection algorithms as well as directional propagations use only the adjacency information of the region adjacency graph. Finally, a complete line-network-extraction system is presented. The performance of the proposed algorithms is studied and a comparison between algorithms operating on the square grid and on the region adjacency graph is made.