SR&ED Case Study: Computer Vision Under Real-World Conditions

Background

A manufacturer’s quality-inspection system used a computer vision model trained and validated under controlled lighting on a single product line. Expanding to a second production line introduced variable lighting, reflective surfaces, and a wider range of defect types.

The Challenge

The existing model’s accuracy dropped sharply under the new conditions, and it was unclear whether the drop could be resolved through data augmentation alone or required a fundamentally different modeling approach.

Technological Uncertainty

It was not known in advance whether any combination of preprocessing, augmentation, and model architecture could maintain acceptable detection accuracy across the variable lighting and reflective-surface conditions of the new line.

Experimental Development

The team systematically tested lighting-normalization preprocessing techniques, synthetic data augmentation strategies, and alternative model architectures, measuring detection accuracy and false-rejection rates for each combination under real production conditions.

What Failed?

Standard data augmentation techniques improved accuracy under simulated lighting variation but did not generalize to the actual reflective surface artifacts seen in production, requiring the team to develop a custom preprocessing step.

Technological Advancement

The team developed a combined preprocessing and model-retraining approach that restored detection accuracy to an acceptable level across both production lines, generating new understanding of how to generalize the vision system across varying physical conditions.

Potentially Relevant SR&ED Activities

Systematic testing of preprocessing and augmentation strategies against real production variability
Architecture experiments to identify a model configuration robust to lighting and surface variation
Accuracy and false-rejection benchmarking across iterations under live conditions

What Would Generally Not Qualify

Deploying the finalized, validated model to additional identical production lines with the same physical conditions would not itself involve further eligible experimental development.

Documentation

Accuracy benchmarks across iterations, preprocessing technique notes, and records of the specific real-world failure modes investigated would support this claim.

About The Author

Dale Doering

Dale Doering is the owner of SRED Consultants Inc., helping businesses navigate the complexities of Scientific Research and Experimental Development (SR&ED) claims. With a strong understanding of the technical and interpretive requirements of the SR&ED program, Dale works with companies to identify eligible projects, document technological challenges, and clearly demonstrate the systematic experimentation or analysis undertaken to achieve advancement. His approach focuses on translating complex technical work into well-supported SR&ED claims, helping clients maximize eligible opportunities while maintaining a clear understanding of the program’s requirements.

Frequently Asked Questions

Why did the initial computer vision model perform poorly on the new production line?

The original model was trained and validated in a controlled environment on a single product line. When introduced to the second line, it encountered real-world physical variations—such as unpredictable lighting, reflective surface artifacts, and a broader variety of defects—which caused a sharp drop in detection accuracy.

What was the main technological uncertainty investigated in this project?

The team faced uncertainty over whether standard adjustments (like data augmentation or minor preprocessing) could resolve the performance drop, or if achieving acceptable accuracy across variable lighting and reflective surfaces would require a fundamentally different model architecture.

Why did standard data augmentation techniques fail to solve the problem?

While standard data augmentation improved performance in simulated lighting conditions, it failed to replicate or generalize to the complex reflective surface artifacts present in the actual production environment. This required the team to create a custom preprocessing step specifically tailored to those real-world conditions.

What key activities from this project would potentially qualify for SR&ED?

Potentially eligible experimental development activities include:

– Systematic testing of preprocessing and augmentation strategies against live production variability.

– Architecture experiments to build a model robust to surface and lighting variations.

– Iterative benchmarking of accuracy and false-rejection rates under real-world conditions.

What part of the deployment process would not qualify for SR&ED tax credits?

Once the combined preprocessing and model-retraining solution was finalized and validated, scaling or deploying that exact model to additional, identical production lines with the same physical conditions would generally not qualify as experimental development.