Quick answer: Installing cameras, labelling images, training an established vision model and connecting it to a production line will often be implementation rather than experimental R&D. A potential core R&D activity may arise where a narrower outcome cannot be known or determined in advance by a competent professional on the basis of current knowledge, information or experience — for example, whether a rare defect can be detected at a defined recall and false-positive rate under the required operating conditions.
The Australian Government has announced R&DTI reforms that will apply to income years starting on or after 1 July 2028. Until those changes apply, activities continue to be assessed under the current rules.
A manufacturer may have a straightforward objective:
Inspect every product automatically and remove defective units before they leave the line.
That is an AI project. It is not necessarily an R&D activity. Commercial cameras, annotation platforms, established vision architectures and edge hardware are readily available. In many cases, a competent professional can use those tools to build a working inspection system with a reasonably predictable outcome.
The harder question appears when defects are rare, visually subtle or easily confused with normal production variation. The relevant uncertainty may then become:
Can the available visual signal distinguish the defined defect from acceptable variation, at the required detection rate and false-alarm level, under the conditions in which the system must operate?
That is a narrower and potentially experimental question.
A 99% Accurate Model Can Still Be Operationally Useless
Rare-defect inspection involves extreme class imbalance. Consider a production run of 100,000 items in which only 0.2% are defective:
• 200 defective items
• 99,800 acceptable items
Suppose a model detects 90% of the defects but incorrectly flags 1% of acceptable items. It finds approximately 180 genuine defects but also produces approximately 998 false alarms. Overall accuracy is close to 99%, yet only around 15% of the items sent for review are actually defective. The model appears accurate while creating an operationally poor inspection process.
For this reason, conventional accuracy — and sometimes ROC-AUC viewed in isolation — may obscure the performance that matters in a severely imbalanced production setting. More informative measures may include:
• Recall by defect class
• Precision and the Precision–Recall Curve or PR-AUC
• An F-beta score weighted towards recall where missing a defect carries greater risk
• False alarms per hour or production shift
• Minimum detectable defect size
• Inference latency
• Performance across specified speeds, finishes and environmental conditions
The metric should match the operating decision. “Can we improve the model?” is too open-ended. “Can defects of the defined class be detected with at least 90% recall while maintaining a false-positive rate below 0.2% at the required line speed?” is testable.
The Inspection System Is Not the Candidate Core Activity
Under the current R&DTI rules, a core R&D activity must involve an outcome that cannot be known or determined in advance and can only be determined through a systematic progression of work based on established science, proceeding from hypothesis through experiment, observation and evaluation to logical conclusions. It must also be conducted for the purpose of generating new knowledge. Much of a machine-vision project may therefore sit outside the candidate core activity:
Usually closer to implementation:
• Selecting and installing commercial cameras
• Configuring standard lighting equipment
• Building routine image-ingestion and annotation pipelines
• Deploying an established model architecture
• Connecting outputs to a PLC, MES or quality system
• Developing dashboards, alerts and reports
• Adjusting familiar parameters where their effects are reasonably understood
• Rolling out a demonstrated method across additional lines
These tasks may require significant engineering skill and expenditure. Difficulty and cost do not establish that the outcome was unknowable.
Where an experimental question may arise:
• Investigating whether a subtle defect produces sufficient detectable signal at operating speed
• Determining whether reliable detection is possible from very few independent defect examples
• Evaluating whether performance survives changes in lighting, reflectivity, contamination or surface texture
• Resolving whether a defect can be separated from acceptable process variation
• Testing whether synthetic defects preserve the physical features required for real-world detection
• Proving whether the required recall can be achieved without an unacceptable false-alarm burden
The distinction is not whether machine learning is involved. It is whether a competent professional could determine the achievable outcome in advance.
Rare Defects Change the Research Problem
A large image count does not necessarily provide a large amount of independent information. If thousands of frames come from the same small group of defective products, a random image-level train/test split may place nearly identical observations on both sides. The reported result may reflect memorisation rather than generalisation. Splitting by physical unit, production batch or capture event may produce a very different result.
The imaging method also matters. If the camera does not capture the feature that distinguishes the defect, further model training cannot recover information that was never recorded. Resolution, exposure, wavelength, illumination geometry, polarisation and motion blur may therefore form part of the investigation.
Synthetic data presents a similar issue. A generated defect may look convincing while failing to reproduce the physical signal found in a genuine failure. Its value should ultimately be evaluated against independent real-world observations.
Ground truth also needs scrutiny. If experienced inspectors regularly disagree about whether an item is defective, model performance against either inspector cannot automatically be treated as objective truth. Useful records may include:
• The defect definition and acceptance criteria
• The annotation protocol
• Independent expert review
• Microscopy, measurement or destructive-test results where relevant
• Disagreements between assessors
• The treatment of ambiguous samples
A small apparent improvement between models may not be meaningful if label uncertainty is greater than the measured difference.
Define the Question Before Training the Model
Consider a hypothetical manufacturer producing coated aluminium components. A rare surface defect appears after processing, but changes in texture and reflection also generate visually similar patterns. The candidate research question might be:
Can an automated vision method detect defects of the defined type and size with recall of at least 90% and a false-positive rate below 0.2%, across the specified surface finishes and at the required operating speed?
The prior-knowledge assessment should then examine published research, vendor specifications, accessible benchmarks and internal experience. If that material allows a competent professional to determine the answer, the work is closer to implementation. If it does not, the team may have a technical uncertainty worth testing. A hypothesis could be:
Controlled multi-angle illumination combined with a defect-localisation model will preserve the relevant defect signal across the specified finishes sufficiently to meet the pre-defined recall and false-positive thresholds.
The team could first test this proposition on an offline rig replicating the relevant production dynamics. Variables may include lighting geometry, polarisation, exposure, resolution, simulated line speed, preprocessing method and model approach. The purpose is not to prove that AI works. It is to determine whether the hypothesis is supported.
A negative result may still generate useful knowledge. The team may establish that the defect signal disappears above a particular speed, that an augmentation method creates feature drift, or that the available imaging modality cannot achieve the required false-positive performance.
Supporting Activities and Production Testing
Activities that are not core R&D may still need to be assessed under the supporting R&D activity rules. For example, configuring an experimental camera rig, preparing a dataset specifically for the experiment or connecting model outputs to test equipment may potentially be supporting activities where they are directly related to a core R&D activity.
Dominant purpose test & production boundaries
Where an activity is excluded from being core R&D, produces goods or services, or is directly related to producing goods or services, it must also be undertaken for the dominant purpose of supporting the core R&D activity. This does not mean that every installation, labelling or integration cost is eligible. The relationship and purpose must be assessed activity by activity.
Testing on an offline rig may make experimental control and activity boundaries clearer, but testing on a commercial production line is not automatically disqualified. Core R&D activities can occur in a production environment if the statutory requirements are otherwise met.
Separate tax issues may nevertheless arise. The feedstock adjustment provisions in s 355-445 should be considered where claimed feedstock inputs are transformed or processed during R&D activities into tangible outputs that are later supplied or applied to the company’s own use. Merely inspecting already manufactured products on a production line does not, by itself, establish that a feedstock adjustment applies. The treatment depends on the actual inputs, transformation process, outputs, expenditure and use of those outputs.
The dominant-purpose test and feedstock provisions should therefore be considered separately rather than treated as a general prohibition on production testing.
What Useful Contemporaneous Evidence Looks Like
Experimental element
Typical engineering record
Stronger contemporaneous R&D record
Prior knowledge
“Reviewed model documentation and camera specifications.”
Records which available models and imaging methods were reviewed, the operating conditions covered, and why they did not establish whether the required performance was achievable.
Technical hypothesis
“Optimise the model to reduce missed defects.”
A dated proposition linking a proposed sensing and modelling method to defined recall, precision or false-alarm thresholds under specified conditions.
Experimentation
Repeated tuning on a combined dataset until a satisfactory result appears.
A controlled trial plan identifying variables, comparison methods, dataset splits, test conditions and evaluation criteria before results are known.
Observation
“Model accuracy reached 99%.”
Results by defect class and operating condition, including recall, precision, PR-AUC, false alarms per shift, latency and observed failure modes.
Evaluation & conclusion
“The model is ready for deployment.”
An assessment of whether the hypothesis was supported, what technical limits were identified, what failed and what changed before the next experiment.
business.gov.au guidance states that records should be created while activities are being conducted. For machine-vision work, useful records may also include dataset manifests, model and preprocessing versions, camera settings, lighting configurations, batch identifiers, ground-truth methods and explanations of how experimental work was separated from routine deployment. Evidence does not convert implementation into R&D. It helps demonstrate what was actually investigated and when.
The Better Question Is Not “Is Computer Vision R&D?”
Computer vision can be routine implementation. It can also sit within a genuine experimental investigation. The technology does not decide which one it is. The more useful starting point is:
Could a competent professional determine in advance, from knowledge, information or experience that was publicly available or reasonably accessible anywhere in the world, that the required performance was achievable using the available data, sensing method and operating conditions?
If yes, the work is likely closer to implementation. If the answer is genuinely unknown, the next step is to define the technical uncertainty, hypothesis, variables and evaluation criteria before beginning undirected model tuning. That is where an automated inspection project may begin to contain research.
Sources & Further Reading
Australian Government, business.gov.au — Conducting core R&D activities for the R&D Tax Incentive
Australian Government, business.gov.au — Conducting supporting R&D activities for the R&D Tax Incentive
Australian Government, business.gov.au — Artificial intelligence-related activities and the R&D Tax Incentive
Australian Government, business.gov.au — Record keeping for the R&D Tax Incentive
Federal Register of Legislation — Income Tax Assessment Act 1997, Division 355, including ss 355-25, 355-30 and 355-445
Australian Taxation Office — Clawback of feedstock adjustments
Saito, T. and Rehmsmeier, M. — The Precision–Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets, PLOS ONE (2015)
Ignition Research — When Does an AI Project Become More Than Implementation?
Ignition Research — Scanning the Site Is Service Delivery: Where the R&D Sits in Turning Point Clouds into Decisions
Australian Government, business.gov.au — 2026–27 Budget: R&D Tax Incentive reforms
Talk to Ignition Research if you are developing machine-vision, automated-inspection or manufacturing technology and need technical R&D support. As a Registered Research Service Provider, Ignition Research assists businesses with research question definition, experimental design, technical investigations and contemporaneous supporting records within its registered research fields.
RSP registration confirms that the provider has met the relevant registration requirements for its registered research fields. It does not constitute government endorsement of a project or guarantee R&DTI eligibility.
We do not determine R&DTI eligibility or provide tax advice. The R&D Tax Incentive is a self-assessment program, and companies remain responsible for assessing their activities, expenditure and circumstances and obtaining appropriate tax advice.
This article provides general information only and is not tax, legal or financial advice.
Working on a project like this?
Tell us what you're building and the part you're not sure will work. A researcher from our Adelaide team (registered Research Service Provider RSP000047) reads every enquiry and replies within two business days. No documents needed.
