Service overview
Machine learning, predictive maintenance, computer vision, sensor analytics and forecasting. This service is part of that division and is scoped individually around what you need to learn, decide or publish.
Images are a rich source of engineering information: cracks on a surface, grain structures in a micrograph, deformation patterns in a test, defects on a production line. Extracting quantitative information from them has traditionally relied on hand-crafted algorithms and expert inspection. Machine learning, particularly deep learning, has expanded what can be automated, but it also introduces new failure modes.
The most important determinant of success is often not the network architecture but the images: how they were acquired, how consistent the lighting and focus are, how the labels were produced and whether the images used for development resemble those the system will meet in use. A model trained on pristine lab images may fail on images from a different camera or lighting condition. This service emphasises acquisition, labelling quality and evaluation under realistic variation.
If you plan to publish the outcome, see manuscript & publication support, for example computational & simulation manuscripts or original research article writing.
What this service covers
Problems this service can address
- Too few labelled images, or labels of uncertain quality
- Models that fail when lighting or camera set-up changes
- Unclear acceptable error for the application
Methods and analytical approaches
- Image pre-processing and classical measurement methods
- Classification, detection and segmentation models
- Data augmentation and annotation-quality review
- Evaluation on independent imaging conditions
Applications
- Surface-defect and crack detection
- Microscopy image quantification
- Displacement and deformation measurement
Relevant engineering disciplines
Scope, inputs and deliverables
Scope of work
- Data audit and definition of the prediction or decision task
- Model development with structure-aware validation and baselines
- Evaluation, error analysis and uncertainty reporting
- Documentation sufficient for others to reproduce the work
- Feasibility check on: representative image set and acquisition details
Typical inputs required from you
- Representative image set and acquisition details
- Definition of the target feature or defect
- Ground-truth annotations, if available
Deliverables
- Analysis pipeline description
- Performance evaluation with failure-case examples
- Notes on image-acquisition improvements
Highly specialised requests are first checked for feasibility from this information. Please do not send confidential data through the public form; describe it first and agree secure transfer afterwards.
Project workflow and method
- 01Enquiry and requirement assessment
- 02Technical scoping and proposal
- 03Execution
- 04Quality review and delivery, with an agreed round of revisions
Read the full project workflow.
In depth: how the work is done
Image acquisition and pre-processing
Image quality limits everything downstream. We review camera, lens, resolution, illumination, mounting and calibration, and recommend changes that make the task easier, such as diffuse lighting to reduce glare or scale references for measurement. Pre-processing steps, including normalisation, denoising, geometric correction and contrast enhancement, are chosen with care and documented, since they influence both classical and learned methods.
Classical measurement methods
Thresholding, edge detection, morphological operations, template matching and feature-based methods remain effective when the imaging conditions are controlled and the targets are well defined. They are transparent, need no labelled training data and are easy to validate. Dimensional measurement from images requires calibration and uncertainty estimation, including lens distortion and pixel-sise effects. We use classical methods where they suffice and as baselines for learned approaches.
Classification, detection and segmentation models
Convolutional and transformer-based models can classify images, locate objects with bounding boxes or delineate regions pixel by pixel. For small engineering datasets, transfer learning from models pretrained on large image collections and strong augmentation are common strategies. Architecture choice is secondary to data quality; we select models proportionate to the data and the required output, and we compare them with simpler baselines.
Annotation quality and dataset design
Labels define what the model learns. Inconsistent annotations, especially for ambiguous boundaries such as crack tips or grain edges, limit performance and make evaluation unreliable. We review annotation guidelines, measure agreement between annotators and refine labels where needed. Datasets are split so that images from the same specimen, session or part do not appear in both training and test sets, avoiding leakage that would overstate performance.
Evaluation under realistic conditions
Performance should be measured on independent data that include the variation expected in use: different cameras, lighting, surface conditions and operators. Metrics appropriate to the task are used, such as intersection over union for segmentation or precision and recall at a chosen threshold for defect detection, and error cases are examined visually to understand failure modes. We report limitations explicitly, including conditions under which the system should not be used.
Quality control and good practice
Quality-control approach
- Data are audited for quality, leakage and drift before modelling.
- Validation respects data structure (by specimen, machine or time) and includes simple baselines.
- Performance is reported with uncertainty and failure cases, not only headline scores.
- Model limits and the conditions under which it should not be trusted are documented.
Common pitfalls this service helps avoid
- Training and testing on images from one session. Models may learn the lighting or background and fail elsewhere.
- Inconsistent labels. Ambiguity in annotation sets a ceiling on measured and real performance.
- Ignoring calibration for measurement. Without calibration, pixel measurements cannot be converted to physical dimensions with known uncertainty.
- Using accuracy for rare defects. Accuracy can be high for a model that never detects the defect when defects are rare.
What good reporting looks like
An image-analysis report describes the acquisition set-up, the pre-processing, the dataset with size, annotation process and splits, the methods compared including classical baselines, training procedure, the evaluation metrics on independent data with examples of successes and failures, measurement calibration and uncertainty where relevant, and limitations. It states the imaging conditions within which results apply and recommends how the system should be monitored in use.
An illustrative scenario
Hypothetical example written for explanation. It is not a client case, a result or a testimonial.
Suppose a research group wants to detect cracks in concrete surface photographs. A first model trained on images from one site performs well in validation, but a test on photographs from another site shows many false alarms from formwork lines and stains. The team adds images from several sites, standardises annotation of what counts as a crack, and evaluates by site. The final report gives performance per site and recommends a human review step for low-confidence detections.
Questions to settle before you begin
- Can I control lighting, focus and viewpoint during acquisition?
- How will labels be defined, and by whom?
- Will the system meet images from new cameras or sites?
- Which error is more costly: missed defects or false alarms?
- Do I need physical measurements, and how will I calibrate them?
Key terms
- Segmentation
- Assigning each pixel to a class, such as crack or background.
- Intersection over union
- A measure of overlap between predicted and true regions.
- Transfer learning
- Starting from a model trained on other data and adapting it.
- Augmentation
- Creating modified training images to increase variety.
- Annotation agreement
- A measure of consistency between different labellers.
Working together on AI, Machine Learning & Data Science projects
AI and data science engagements start with the decision the model is meant to support and with an audit of the data: sensors, sampling, labels, missing values, the structure of the data by specimen, machine or time, and how future data will differ from past data. We ask for a representative sample and a description of the process. Validation is designed around that structure to avoid leakage, and results are compared with simple baselines. Reports give performance with uncertainty, failure cases and the conditions under which the model should not be trusted. No performance figures are promised in advance.
Frequently asked questions
Can you build a production inspection system?
Our scope is research-grade analysis and documentation. Production deployment and integration are outside the service scope.
How many images do I need?
It depends on task difficulty and variation. Transfer learning reduces the need, but diversity matters more than volume. A pilot indicates feasibility.
Can you analyse microscope images?
Yes. Quantification of features such as grains, pores or phases is common, with attention to sample preparation and scale calibration.
Can the system run in real time?
That depends on model size and hardware. Deployment is outside the standard scope but can be discussed.
Do you label the images for me?
Annotation guidelines and review can be provided. Domain-expert labelling is best done by people who understand the material or defect.
How is this service scoped and priced?
After your enquiry we review what you need, ask for missing information and send a written scope and quotation. Price depends on complexity, deliverables and timeline. See the project workflow.