Abstract of the Offer
This micro company is located in Cambridge, UK. It is offering ImageTriage, an extensible software platform for analysing disorganised and haphazard images. ImageTriage can be used for automatic processing of NDT/SHM data and generating draft inspection reports across energy and construction industires, manufacturing, etc. It emplys Decision Trees and fuzzy logic rather than Deep Learning AI, making it data lean, agile, explanatory and energy efficient. Seeking partners for validation pilots.
Description
Given A-scans the current version of the software it can employ signal processing functions for surface profiling, meshing, ray tracing and intensity generator to generate standard two-dimensional TFM (Total Focusing Method) or sectorial images or else a series of images, which utilise different portions of the PAUT (Phased Array of Ultrasonic Transducers) or else a combination of scans secured with ordinary probes. These images are processed by image processing functions, whichemploy fast algorithms from the OpenCV (Open-Source Computer Vision Library).
The AI section of ImageTriage contains Decision Trees, which select images that appear to contain defects, evaluates their characteristics, groups similar images and produces a draft report. This form of AI has been chosen, because
- Decision Trees can be trained and retrained on much fewer datasets than neural networks, and therefore are data-lean;
- Decision Trees do not require costly GPU hardware, can run on ordinary computers and therefore are energy efficient, leaving small carbon print;
- Decision Trees are based on “if… then" rules and fuzzy logic, therefore producing explainable results,
The software was originally trained on data collected with linear PAUTs (Phased Array of Ultrasonic Transducers).
The current capabilities of DDC (Defect Detection and Characterisation) module are as follows:
- using A-scans (presented in.txt, .csv, .dat, .png or .h5 format) is can form 2D cross-sectional images of the inspected component;
- even in presence of realistic noise, it can detect and characterise surface-breaking and subsurface cracks in homogeneous metals; the types of cracks used in training include fatigue, rough-surface, stress-corrosion, and weld-adjacent varieties;
- using various cross-sectional images, it can locate and characterise honeycombing in cement;
- component shapes: rectangular, with surfaces plane or undulated;
- measurement techniques: PAUT (FMC/TFM data), PAUT (sectorial data), any combination of two or three wave modes L and T, allowing to deal with reflections and mode conversions.
- generating draft inspection reports.
The DDC capabilities are easily extendable to:
- characterising other types of damage, reading other input file formats;
- using data from 3D scans (e.g. in .3mf format) and CAD files;
- measurement techniques: conventional mechanical scanning by single-probes in pulse-echo & pitch-catch modes; eddy current; radiography; thermography; optics.
The current capabilities of FFS_ASSESS (vessels and pipes Fitness For Service ASSESSment module):
- using A-scans to create corrosion maps, it can apply ASME & BS/EN standards to make assessments;
- component shapes: rectangular and cylindrical, with perfect surfaces;
- generating draft inspection reports.
The FFS_ASSESS capabilities are easily extendable to
- component shapes: rectangular and cylindrical, undulated surfaces.
- measurement techniques: any techniques allowing to create corrosion maps.
The software offers significant potential within the broad NDT/SHM (Non-Destructive Testing/Structural Health Monitoring) market across energy, construction, manufacturing and airspace sectors. Although first developed for PAUTs, the core technology is modality-agnostic and can analyse any inspection data presented as images. This broad applicability allows the system to support a wide range of inspection scenarios without major redesign.
Advantages and Innovations
The core innovation of ImageTriage lies in its purpose-built AI framework for automated NDT data processing.Rather than interpreting traditional A-scans, ImageTriage focuses on direct analysis of damage images, enabling faster and more intuitive defect identification. Instead of relying on heavy, general-purpose models such as convolutional neural networks, ImageTriage uses custom decision trees tailored to specific inspection tasks. This design delivers several benefits:
- Minimal data requirements: no need for large, fully labelled datasets to train or retrain the system.
- Low computational load: runs efficiently on standard hardware without specialised GPUs.
- Energy efficiency: supports sustainable, lightweight deployment.
- The shift to image-based analysis is deliberate: fast, high-quality algorithms from OpenCV enable precise, low-overhead processing without the complexity of deep-learning pipelines.
- Agile: allows for fast deployment and retraining.
- Explainable AI: decision trees follow clear, traceable logic, so inspectors can understand why a decision was made.
Together, these choices create an AI solution that is transparent, practical, and aligned with real-world NDT workflows, avoiding the cost and opacity typical of general-purpose machine learning.