20  Deployment, monitoring and human-in-the-loop for learning-based vision

Chapter status: Structural placeholder.

This chapter currently contains the approved heading structure but no reviewed technical content.

20.1 From trained model to production system

20.1.1 Model export and interchange (ONNX)

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.1.2 Quantisation, pruning and operator fusion

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.1.3 Runtime targets: CPU, GPU, edge accelerators

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.1.4 Revalidation of the deployed artefact

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.2 Dataset governance

20.2.1 Collection planning against the operating envelope

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.2.2 Labelling protocols and inter-annotator agreement

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.2.3 Curation, exclusion records and provenance

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.2.4 Versioning and model-to-data traceability

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.3 Post-deployment monitoring

20.3.1 Domain shift and concept drift

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.3.2 Statistical drift detectors

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.3.3 Retraining loops and change control

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4 Human-in-the-loop systems

20.4.1 Why humans remain in critical and deep-learning loops

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4.2 Review triggers: low confidence, drift, novelty, borderline cases

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4.3 Interface and workflow design for human review

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4.4 Human-in-the-loop as a labelling and improvement engine

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4.5 Levels of autonomy and human oversight

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.4.6 Automation bias and the limits of human checking

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.5 Explainability and interpretability

20.5.1 Saliency and Grad-CAM

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.5.2 LIME and local surrogates

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.5.3 SHAP and Shapley attributions

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.5.4 Limits of post-hoc explanation

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.5.5 Explanation as debugging versus explanation for audit

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.6 Bias and fairness in industrial vision

20.6.1 Subgroups in industrial data: cavities, suppliers, shifts

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.6.2 Subgroup analysis and reporting

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.6.3 Mitigations: data, loss, imaging

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.

20.6.4 Fairness when the system observes people

Content status: Placeholder.

This subsection has not yet been drafted or technically reviewed.