Tactile Imaging Using Watershed-Based Image Segmentation
Sergio E. Hernandez, Kenneth E. Barner · 2000 · Proceedings of the Fourth International ACM Conference on Assistive Technologies (Assets '00) · doi:10.1145/354324.354332
Summary
This paper from the University of Delaware proposes a new image segmentation method for automatically converting visual images into tactile representations suitable for display on tactile devices such as microcapsule (swell) paper. The core problem addressed is that existing approaches to tactile image creation, particularly multiresolution (MR) edge detection techniques, produce results with significant drawbacks — false boundaries, small negligible regions, and representations that can mislead tactile readers about the actual content of the image. The proposed method uses a watershed-based algorithm as its foundation, treating the gradient magnitude of an image as a geographical surface where water flows downhill to local minima, forming catchment basins that correspond to segmented regions. Because watershed algorithms inherently produce oversegmentation (hundreds or thousands of tiny regions), the paper's key contribution is a novel joint region merging criterion that combines two previously independent approaches: region homogeneity (merging regions with similar intensity values) and edge integrity (preserving boundaries where strong gradient edges exist). These criteria are combined through a weighted matrix using a parameter alpha that dynamically shifts from pure edge integrity (to first eliminate false background regions) toward region homogeneity (to preserve small important regions as the count decreases). The method is built on the TACTICS (TACTile Image Creation System), which simplifies images using edge detection, binary segmentation, and their combination for output on microcapsule paper.
Key findings
The joint region merging criterion produces significantly better segmentations than either homogeneity-based or edge-based merging applied independently. When region homogeneity is used alone (alpha=1), small regions with similar mean values to adjacent regions are merged, but this leaves false boundaries from undesirable background regions. When edge integrity is used alone (alpha=0), weak boundaries are eliminated but small regions are lost even when they represent important image content. The combined criterion with dynamic alpha assignment avoids both problems. The proposed nonlinear pre-filtering and post-filtering using median filters effectively reduces oversegmentation while preserving region boundaries — reducing initial regions by an average of 75% before the merging stage. Compared to multiresolution edge detection, the segmentation approach offers two key advantages for tactile imaging: segmentation results displayed as region edges provide clearer representations of basic image information than MR edge detection at the lowest resolution level; and intermediate segmentation stages do not mislead the tactile reader in the way that intermediate edge detection results can, where partial boundaries suggest false contours. Experimental results on Space Shuttle, Ebola virus, and Moon images demonstrate that the final segmentations contain the main object boundaries as closed contours with small negligible features eliminated.
Relevance
This research addresses a fundamental accessibility challenge: making visual images comprehensible to blind and visually impaired users through automatic tactile conversion. While the specific algorithms are technically detailed, the practical goal is enabling blind users to access the same visual information available to sighted users without requiring manual creation of each tactile graphic by a human transcriber. The watershed segmentation approach produces cleaner, more interpretable tactile images than edge detection alone, which is important because false or ambiguous boundaries on tactile paper can significantly mislead a blind reader who has no visual reference to resolve ambiguities. For modern accessibility practice, this work laid groundwork for automated image-to-tactile conversion pipelines and connects to current efforts in generating tactile graphics from digital images for 3D printing, refreshable braille displays, and other tactile output technologies. The principle that automated image simplification must prioritize meaningful content boundaries over raw detail remains central to accessible image representation.
Tags: tactile graphics · image processing · blindness and low vision · image accessibility · computer vision · non-visual access · accessible graphics