Computer Vision

Computer vision is the field of artificial intelligence that interprets images and video. The National Science Foundation's Computer Vision program described its emphasis as "image representation and interpretation for systems designed to infer properties of the environment from imaging data" (NSF 03-602). Its listed research topics included "Recognition, classification, and identification of objects, people, events, and activities" and "Methods for grouping, comparing, matching, indexing, and retrieving visual data."

Techniques

NIST describes convolutional neural networks as "primarily used for processing grid-like data, such as images" (NIST AI 100-2e2025); see Deep Learning. Reading text from images is a specialized task covered under Optical Character Recognition.

Extracting features from maps

USGS, DARPA, NASA's Jet Propulsion Laboratory, and MITRE ran "a 12-week machine learning competition aimed at accelerating development of AI tools for critical mineral assessments." One of its two challenges was "automated feature extraction from historical maps" (USGS, "Extracting data from maps: Lessons learned from the artificial intelligence for critical mineral assessment competition," 2025). The measured results:

• "Prompt-based extraction (i.e., with user input) of polygons, polylines, and points from geologic maps yielded median F1-scores of 0.77, 0.56, 0.35, respectively."
• USGS noted that "Geologic maps pose numerous challenges for AI workflows because they vary significantly."
• The competition led to "new AI tools to semi-automate key, time-consuming parts of the assessment workflow."

The scores are medians, so half the results fell below them. The other challenge, georeferencing, is covered under Geospatial AI.

Vulnerabilities

NIST AI 100-2e2025 reports that "in PredAI computer vision applications for object detection and classification, well-known cases of adversarial perturbations of input images have caused autonomous vehicles to swerve into lanes going in the opposite direction" and "stop signs to be misclassified as speed limit signs." Small, deliberate changes to an image can change what the model sees.

Why it matters for land and mining claim records

In the USGS results, points were the hardest features to extract and polygons the easiest, and the extraction was prompt-based, with user input. Claim corners, monuments, and section corners on a plat are points. Any location or boundary a vision tool pulls from a map is a lead to be checked against the source map and the recorded description, not a finished record.