Inputs: Preparing the Inspection Record

Begin with a task, not a camera

“Inspect this container” is too broad to define a reliable automated task. Reading an identifier, documenting exterior damage and locating an unusual region in a radiograph involve different evidence. State what the system is expected to observe, which images it can use and what would count as an insufficient input. This makes later performance claims testable rather than merely persuasive.

An optical photograph records reflected light from visible surfaces. A radiographic image records information related to attenuation through the scanned load. Neither is a complete description of the cargo. A seal photograph cannot establish the contents behind a door, while a projection image can combine multiple objects into an ambiguous pattern. The acquisition method places a boundary around every subsequent inference.

Connect identity without assuming it

A useful input record links the image to a capture event and the intended shipment. Record the source identifier, time basis, capture location, equipment and any relevant operator notes. Where several images belong to one inspection, preserve their sequence and relationships. Time stamps from unsynchronized devices should not be treated as a reliable ordering mechanism without an additional check.

ISO 6346 provides a framework for identifying freight containers, including a check digit. It can help validate the structure of a transcribed identifier, but it does not prove that an image belongs to a particular booking or that a displayed identifier is genuine. Compare the image association with the relevant operational record. An OCR correction should be recorded as a correction, not silently substituted into the original output.

Preserve the source and explain transformations

Keep the original acquisition file under appropriate access controls. Cropping, contrast adjustment, compression and format conversion can affect what a model or reviewer sees. Derived images should refer back to the original and record the transformation applied. A reviewer must be able to distinguish information captured by the device from a visualization produced later.

A cryptographic hash can help detect whether a particular file has changed. It cannot establish that the source was accurate, that the depicted container was the intended one or that the capture process was lawful. Provenance therefore includes people, equipment and procedures as well as file integrity. Avoid treating a technical integrity check as a complete chain of custody.

Make poor inputs an explicit outcome

Glare, dirt, motion blur, occlusion and low contrast can undermine optical inspection. Radiographic quality depends on different factors, including acquisition settings, material overlap and equipment behavior. Quality checks should be specific to the task and modality. A file that opens successfully is not necessarily suitable evidence.

Define an “insufficient evidence” path before deployment. It might lead to an authorized recapture, a different view or specialist review. It should not default to a negative finding. Record why the image failed the quality check, because recurring capture problems may reveal a correctable process issue rather than a model weakness.

Keep declarations separate from observations

A manifest, packing list or operator description supplies context, not visual ground truth. Preserve which facts came from documents and which were observed or independently verified. Otherwise a training set can reward a model for reproducing a declared category even when the image cannot support it. That creates a misleading impression of visual understanding.

Reference labels also need provenance. State who assigned them, what supporting examination was available and how disagreements were handled. An inspection outcome may remain uncertain even after review. Retaining uncertainty is more useful than forcing every record into an apparently definitive category.

Respect the people and locations in the record

Images may contain faces, vehicle registrations, customer information and commercially sensitive arrangements. Apply data minimization, access restrictions and a justified retention schedule. European operations should assess applicable GDPR obligations and local rules rather than assume that all port imagery is anonymous industrial data.

Conditions vary between locations such as the Port of Helsinki, Port of Rotterdam, Port of Antwerp-Bruges and Port of Marseille Fos. Lighting, equipment, cargo mix and work practices can differ materially. A documented acquisition record makes those differences visible and gives later evaluation a meaningful starting point.

Reference points

• Bureau International des Containers — container identification and ISO 6346 context.
• World Customs Organization — customs and non-intrusive inspection guidance.
• General Data Protection Regulation — European personal-data requirements.