Abstract
A scanned well log and a scanned topographic map pose the same computational question: how do you turn a grid of ink pixels back into the ordered list of coordinates that the document originally encoded? Cartography has been answering that question for six decades, far longer than anyone has tried to digitise paper well logs at scale, and its answers are unusually transferable because the underlying object is the same in both domains, a thin curvilinear feature drawn on a noisy background and meant to be read as geometry rather than as a picture. This survey traces the published cartographic raster to vector lineage from the parameter space line voting of the early 1960s [1] [2], through the morphological thinning that defined the centreline step [3] [4], to the convolutional segmentation that the map field adopted in the late 2010s [7] [8], and reads it against the canonical map processing survey that catalogued the whole pipeline [6]. We organise the prior art into the five stage pipeline the field converged on, scan, clean, segment, trace, and resample, and align it stage for stage with the well log digitisation pipeline we built. The two pipelines share an output target, a curve resampled to a fixed grid of 300 depth points, and they share a hardest step, the mask to centreline geometry where a pixel mask must become a one pixel wide vector. We read the synthesis against our own three class raster log baseline, where line tracing accuracy on the thin curve masks peaks at an intersection over union of 0.51 against a background mask that scores 0.94, and we argue that the cartographic literature predicts both the shape of the pipeline and the location of that hardest step. The survey's contribution is the crosswalk: cartography is the closest published analog to well log curve extraction, and reading it saves the log digitisation field from rediscovering the pipeline by trial and error.
Background and the shape of the cartographic literature
The raster to vector problem in cartography is older than the field that named it. The instinct to recover lines from a pixel grid by accumulating evidence in a transformed space arrived with the Hough patent in 1962, which proposed voting for the parameters of a line so that scattered, broken pixels along the same straight feature reinforce a single peak [1]. The scheme was elegant but awkward to use until Duda and Hart recast it a decade later in the rho theta polar parameterisation that handles vertical lines gracefully and made the transform practical for real images [2]. That pair of works is where the lineage starts, and it already contains the central idea every later method inherits: a curve in image space is a structured object whose parameters can be recovered, not a loose collection of dark pixels.
The second movement in the literature was morphological. Once a binary mask of the line work exists, the question becomes how to reduce a stroke several pixels wide to the single coordinate string a vector format wants. The answer the field standardised on is thinning, and the Zhang and Suen two subiteration algorithm became the reference implementation, peeling boundary pixels symmetrically until only a one pixel wide skeleton remains while preserving connectivity and endpoints [3]. Thinning is not free of controversy, and the graphics recognition community argued the point directly: Tombre and colleagues asked whether one should thin at all before vectorising, since thinning can introduce spurious branches and distort junctions, and proposed contour based alternatives that avoid the skeleton entirely [4]. That debate matters here because it names the exact stage where accuracy is won or lost, the conversion from a filled mask to a centreline, which is the same stage that turns out to be hardest in our own log work.
The third movement was the assembly of these pieces into end to end map processing systems, and the canonical reference is the Chiang, Leyk, and Knoblock survey, which catalogued how the field separates cartographic colour layers, strips the graticule and text furniture, segments the line features, and traces them to vectors [6]. The same group's pattern tracing pipeline for United States Geological Survey maps showed that road vectorisation could be made to work on heterogeneous scanned sheets with minimal user intervention by exploiting the common geometric properties of the lines [5]. By the time the field turned to learned segmentation, replacing hand tuned colour separation and morphology with convolutional encoder decoders trained to produce the line mask directly [7] [8], the pipeline shape was already fixed. Deep learning changed how the segment stage is implemented, not the order of the stages. That stability is precisely what makes the cartographic literature a usable map for an adjacent domain.
How the crosswalk was assembled
The synthesis here is a structured reading of the published cartographic raster to vector literature, not a new cartographic experiment, and it is worth stating how the crosswalk was built so the reader can judge what it does and does not establish. We took the pipeline that the map processing survey treats as canonical [6] and decomposed it into five stages that recur across the prior art: scan the sheet to a raster, clean off the furniture that is not the target line work, segment the line pixels into a mask, trace that mask to a centreline, and resample the centreline to ordered vector nodes. For each stage we traced the defining contributions back to their originating works, the Hough and Duda and Hart line voting for the early segment instinct [1] [2], the Zhang and Suen thinning and the Tombre vectorization debate for the trace stage [3] [4], the pattern tracing pipeline and the convolutional turn for the modern segment and assembly stages [5] [7] [8].
We then aligned each cartographic stage with its twin in the well log digitisation pipeline we built. The alignment is not metaphorical. A scanned log strip is a greyscale raster with one input channel and variable page dimensions, exactly like a scanned map sheet. The log carries a depth grid, a header, and per track gridlines that must be removed before the curve can be isolated, exactly the furniture the cartographic clean stage strips. The curve pixels are segmented into a per class mask, background plus two curve classes per frame in our multiclass setting, exactly the segment stage. The curve mask is thinned to a centreline before values can be read off it, exactly the trace stage and exactly the step the Tombre debate flags as fragile [4]. And the centreline is read off at a fixed grid of depth steps to produce the final vector, exactly the resample stage. To keep the crosswalk grounded rather than abstract, we read it against a concrete reference point from our own engagement, a three class raster log segmentation problem on public scanned logs from the Texas Railroad Commission archive, where the numbers we quote are real and sourced. The interactive crosswalk below is built on that basis, with real anchor numbers and an illustrative response shape where a per threshold curve was not measured.
Results: a stage for stage crosswalk, read against a real baseline
The synthesised picture has a clear top level shape, and it is the same shape the map processing survey describes [6]: five stages, executed in order, with the difficulty concentrated almost entirely in the conversion from mask to centreline. The interactive crosswalk lets the reader walk the two pipelines side by side and see where they agree and where they diverge.
Four of the five stages transfer almost without friction. Scanning, furniture removal, and final resampling are nearly identical operations whether the page is a map or a log, which is why the cartographic literature on graticule and text suppression reads as direct advice for log header and gridline suppression. The segment stage transfers in form, since both domains now produce a pixel mask with a convolutional encoder decoder of the kind the map field adopted [7] [8], but it is where the two domains' difficulty profiles begin to separate, because a well log curve is thinner and the class imbalance more extreme than a road on a topographic sheet. The shared output target is unambiguous: both pipelines exist to produce one curve resampled to 300 depth points, the fixed grid our validation reads values off, which is the cartographic resample stage under a different name.
The divergence lives entirely in the trace stage, and our baseline makes the cost vivid. On the three class set, the background mask is essentially solved, scoring an intersection over union of 0.94, because background is enormous and easy. The two thin curve masks collapse to an intersection over union of 0.26 and 0.21, and line tracing accuracy on the curves peaks at 0.51. That gap between 0.94 on the easy class and 0.51 at the curve tracing peak is the entire story of the engagement compressed into two numbers, and it is exactly the gap the Tombre debate predicted [4]: the hard part was never finding roughly where the line is, it was committing to a single one pixel wide centreline through a mask that is fuzzy at its edges and occasionally broken along its length. Where the mask to centreline geometry is clean, downstream regression of the curve values is excellent, with a peak coefficient of determination of 0.9891 on a recovered curve, which says the resample stage does its job once the trace stage hands it a faithful centreline. The accuracy ceiling is set upstream, at the trace, precisely where the cartographic literature located it forty years ago.
Discussion: why cartography is the right analog, and where our work sits
Laid out as a crosswalk, the cartographic prior art stops being a list of old papers and becomes a design map for an adjacent problem. The reason the transfer is so clean is that the two domains share an object and a goal. In both, the thing on the page is a thin curvilinear feature meant to be read as geometry, and the goal is to recover the ordered coordinates that geometry encodes. Document image analysis writ large is the broad neighbourhood, but cartography is the specific house, because maps share with logs the extreme aspect ratios, the dense overprinted furniture, the heterogeneous scan quality across an archive, and above all the requirement that the output be a faithful vector and not merely a pretty mask. A road extraction system that vectorises a United States Geological Survey sheet with minimal user input [5] is solving a structurally identical task to a system that vectorises a scanned log strip, and it is solving it on the harder, older, better studied side of the problem.
Where our own work sits in this landscape is worth stating plainly, because it marks the boundary between this survey and our build. This piece is a reading of the public cartographic field; the well log digitisation system we describe, including the synthetic data pipeline and the segmentation and centreline models that produced the baseline numbers, is ours, and we call it VeerNet. The value of the survey is not that it replaces that build but that it explains it. The cartographic literature predicted the pipeline shape we converged on, so our pipeline is not idiosyncratic, it is the same five stages the map field settled on decades earlier. It predicted the location of the hardest step, so the disproportionate effort we spent on the mask to centreline conversion was not a local accident, it was the field's known bottleneck arriving on schedule. And it predicted the trade we faced at that step, thin or not to thin, mask fidelity against centreline stability, which is the Tombre question restated for a one to three pixel curve [4]. A practitioner who reads the cartographic raster to vector lineage before building a log digitiser will recognise their own roadmap in it, and will know to budget for the trace stage rather than the segment stage.
The practical reading is therefore not to memorise a method but to inherit a map. If the target is a thin curvilinear feature on a scanned page, the cartographic pipeline is the prior to start from, the trace stage is where to concentrate engineering, and the resample target, a fixed grid of coordinates, is the right way to define done. Cartography did the hard generalisation work first; the log digitisation field can stand on it instead of repeating it.
Limitations
This is a survey, and it carries the limits of one. It synthesises what the published cartographic literature reports and aligns that literature with our own pipeline; it does not re-run the cartographic methods on log data, nor does it benchmark a map vectoriser against a log vectoriser under a common protocol, so the crosswalk is an argument from structural correspondence rather than from a head to head measurement. Where the survey quotes numbers, the 300 point resample target, the per class intersection over union of 0.94, 0.26, and 0.21, the 0.51 line tracing peak, and the 0.9891 peak coefficient of determination are real and drawn from a single engagement, a three class problem on one public scanned archive, not from a fresh multi domain study, and the relative difficulty of the trace stage observed there may compress on thicker features or milder imbalance. The recovered fraction rendered against the overlap floor in the interactive crosswalk is an illustrative monotone response anchored on the real 0.51 peak, not a per threshold measured curve. The cartographic works surveyed are a representative spine of the lineage, the line voting, the thinning, the vectorization debate, the canonical survey, and the convolutional turn, not an exhaustive bibliography of map processing, and the analogy holds most strongly for line and curve features and weakest for area and text features that have no clean well log counterpart. A reader should treat the crosswalk as a guide to where the cartographic literature is worth mining and where the two problems genuinely differ, not as a claim that a map vectoriser will run unmodified on a well log.
Key takeaways
- Cartography has solved the raster to vector problem for six decades, far longer than anyone has digitised paper well logs at scale, which makes it the closest published analog to pulling a curve off a scanned log: the same object, a thin curvilinear feature, and the same goal, recover the ordered coordinates it encodes.
- The cartographic literature converged on a five stage pipeline (scan, clean, segment, trace, resample) that maps stage for stage onto well log digitisation. Four stages transfer almost without friction; the divergence is concentrated entirely in the trace stage, the conversion from a pixel mask to a one pixel wide centreline.
- Both pipelines share one output target: a curve resampled to a fixed grid of 300 depth points, which is the cartographic resample stage under a different name. Once the trace stage hands over a faithful centreline, downstream value regression is excellent, with a peak coefficient of determination of 0.9891.
- Our real three class baseline locates the bottleneck exactly where the cartographic literature predicted: the background mask scores an intersection over union of 0.94 while the thin curve masks fall to 0.26 and 0.21 and line tracing accuracy peaks at 0.51. The thin or not to thin debate from graphics recognition named this fragile mask to centreline step decades ago.
- This is a survey of the public cartographic field; the log digitisation system, VeerNet, is ours. The survey does not replace the build, it explains it: cartography predicted the pipeline shape, the location of the hardest step, and the trade at that step, so a team can inherit the roadmap and budget for the trace stage rather than rediscover it.
References
[1] Hough, P. V. C. Method and Means for Recognizing Complex Patterns. U.S. Patent 3,069,654 (1962). The parameter space voting scheme for detecting lines in images that opens the raster to vector lineage. https://patents.google.com/patent/US3069654A/en
[2] Duda, R. O., and Hart, P. E. Use of the Hough Transformation to Detect Lines and Curves in Pictures. Communications of the ACM, 15(1) (1972). The rho theta polar parameterisation that made the Hough transform practical for line extraction. https://dl.acm.org/doi/10.1145/361237.361242
[3] Zhang, T. Y., and Suen, C. Y. A Fast Parallel Algorithm for Thinning Digital Patterns. Communications of the ACM, 27(3) (1984). The two subiteration thinning that reduces a line mask to a one pixel wide skeleton, the canonical centreline step. https://dl.acm.org/doi/10.1145/357994.358023
[4] Tombre, K., Ah-Soon, C., Dosch, P., Masini, G., and Tabbone, S. Stable and Robust Vectorization: To Thin or not to Thin. Graphics Recognition (GREC), Springer LNCS (2000). The vectorization debate over whether to thin a line mask before tracing it, framing the trade the centreline step lives in. https://members.loria.fr/KTombre/tombre-icpr00.pdf
[5] Chiang, Y.-Y., and Knoblock, C. A. A general approach for extracting road vector data from raster maps. International Journal on Document Analysis and Recognition (IJDAR), 16 (2013). A parallel pattern tracing pipeline that vectorises roads from heterogeneous scanned maps with minimal user input. https://link.springer.com/article/10.1007/s10032-011-0177-1
[6] Chiang, Y.-Y., Leyk, S., and Knoblock, C. A. A Survey of Digital Map Processing Techniques. ACM Computing Surveys, 47(1) (2014). The canonical survey of how the field separates colour layers, removes furniture, and vectorises line features from scanned maps. https://dl.acm.org/doi/10.1145/2557423
[7] Ronneberger, O., Fischer, P., and Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI (2015). The symmetric encoder decoder with skip connections that learns dense masks from scarce labels, the segmentation backbone the map field adopted. https://arxiv.org/abs/1505.04597
[8] Uhl, J. H., Leyk, S., Chiang, Y.-Y., Duan, W., and Knoblock, C. A. Automating Information Extraction from Large Historical Topographic Map Archives: New Opportunities and Challenges. In Handbook of Big Geospatial Data, Springer (2020). The shift to convolutional segmentation for recognising features across large scanned map archives. https://link.springer.com/chapter/10.1007/978-3-030-55462-0_20
