Corrosion Detection Image Datasets for Asset Integrity Inspection

What public corrosion and rust image datasets actually cover

Public data gives you a starting point for rust as a class, but rarely for steel-asset severity grading. Civil infrastructure drone research labels rust alongside cracks, and the defect classes are small relative to background, which is why work on that imagery focuses on active learning for imbalance [1]. Those labels typically say "rust present," not how much coated carbon steel has blistered, undercut or pitted. Commercial annotation vendors list corrosion as a standard label in drone inspection projects [2], which tells you the class is routine, not that graded, asset-specific data is available off the shelf.

Broader defect benchmarks vary in realism. A review of manufacturing defect benchmarks notes that DAGM is synthetic and carries at most one defect per image [3], a reminder to check whether a candidate set reflects field conditions at all. For the general trade-offs, see the industrial defect image datasets guide and the image data hub.

Severity scales: pick one before you buy

Severity labels are only comparable when every image was graded against the same written scale. For painted steel, published rust-grading practices rate the share of surface rusted against written definitions and photographic references; ask suppliers which standard and edition they used, and buy the standard text yourself so your annotators grade against the same definitions. Percentage-of-surface scales do not describe rust creep from a damage point, so undercutting needs its own label.

Many operators grade with internal schemes instead, such as A/B/C remediation priorities or "monitor, plan, repair now" bands. Those are usable if the supplier gives you the definitions, the version in force when each image was graded, and a mapping to your target taxonomy. Ask who graded: a certified coatings or corrosion inspector, a field technician, or an offshore annotation team working from photos alone.

Failure modes to look for:

• Scale drift: the grading guide changed mid-collection and nobody versioned it.
• Staining vs. section loss: surface rust bloom labeled the same as pitting with measurable wall loss.
• Viewpoint bias: close-ups of the worst spot, with no wide shot showing the share of surface affected, which makes percentage-based scales impossible to verify.
• Coating confusion: chalking, discoloration, bird droppings or wet mill scale mislabeled as rust.

Asset coverage and capture conditions

A corrosion model generalizes only across the asset types, coatings and capture platforms it has seen. Specify coverage explicitly: pipe racks and insulated piping (corrosion under insulation often shows only as damaged cladding or staining), storage tank shells and roofs, offshore splash zones, ship hulls and ballast tanks, transmission towers and substation steel, bridges and pressure vessel exteriors.

Capture platform matters as much as asset. Drone RGB from 10 to 30 meters, rope-access handheld close-ups, crawler cameras and fixed cameras produce different scales and blur. Thermal and infrared imagery can flag moisture under insulation; for that modality see thermal and infrared inspection images. Ask for camera model, focal length, GSD or distance, and lighting, and decide what to do with EXIF GPS before delivery, as covered in EXIF metadata in image training data.

Linking images to inspection reports and work orders

The most valuable corrosion data connects each photo to what the inspector wrote and what maintenance did next. An image tagged "grade 6" is useful; the same image linked to a wall-thickness reading, a coating repair work order, and a follow-up photo after blasting and recoat lets you train severity models against outcomes, not just opinions. Inspection reports and maintenance logs are often the richer asset; see the SourceX pages to license inspection reports and license maintenance logs.

Paired pre- and post-remediation photos also support change detection; the before-and-after image pairs guide covers alignment and pairing keys. Free-text inspector notes can become domain captions, as described in domain captions from work records.

Request template for a corrosion image dataset

A precise request separates what the model needs from what is nice to have. Use the record below as a starting schema when you describe data to suppliers or brokers, and ask for a Croissant JSON-LD descriptor so the file and field structure is machine-readable [4].

Illustrative example: invented to show structure; it does not describe an available dataset.

{
  "image_id": "img_000123",
  "asset_type": "storage_tank_shell",
  "component": "shell_course_2_north",
  "substrate": "carbon_steel",
  "coating_system": "epoxy_polyurethane_3coat",
  "capture": {"platform": "drone_rgb", "distance_m": 12, "camera": "redacted_model", "date": "2024-06"},
  "labels": {
    "geometry": "polygon_mask",
    "classes": ["general_rust", "blistering", "undercutting_at_damage"],
    "severity_scale": "owner_scale_v3",
    "severity_grade": "priority_B",
    "grader_role": "certified_coatings_inspector",
    "grading_guide_version": "v3"
  },
  "linked_records": {
    "inspection_report_id": "IR-2024-0412",
    "ut_min_thickness_mm": 7.9,
    "work_order_id": "WO-55821",
    "remediation": "spot_blast_and_recoat",
    "post_repair_image_id": "img_004410"
  },
  "redaction": {"exif_gps": "removed", "faces": "blurred", "site_signage": "blurred"}
}

Acceptance checklist for a sample:

Rare severe grades are the usual gap; options for pooling or synthesizing them are compared in rare defect coverage, and broader acceptance testing is covered in the training data quality guide.

Rights, people and site sensitivity in inspection imagery

Inspection photos belong to the asset owner or the inspection contractor, and the contract between them decides who can license the images for AI training. Confirm which party holds the rights, whether client confidentiality clauses cover the photos, and whether site security rules restrict imagery of refineries, substations or port facilities.

Field photos also capture workers, vehicle plates, company signage and GPS tags. Ask for a documented redaction method and a checked sample; the face data consent and anonymization guide and property releases for training images cover the details.

How SourceX sources corrosion inspection imagery

SourceX sources operational datasets from US companies on request, including new recordings of hands-on work, and manages the commercial process through licensing and ongoing purchases. Nothing is held in stock and a request does not guarantee a match; buyers describe the data, and every release is approved by the supplying company. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery, with personal details removed or replaced, the method recorded and a sample checked. You can start by describing your target scale, asset types and linked records on the SourceX buyer page, or review the page to license images and inspection photos.

Request corrosion and coating-failure images for your model

SourceX looks for US businesses that hold the corrosion imagery, inspection reports and work orders you describe, then works through assessment of data and licensing permissions before any agreement. Pricing and allowed uses are agreed per deal in a license, and delivery runs through private, access-controlled workflows after an executed agreement. Describe the corrosion image data you need.