Industrial vision · Feasibility study
A lightweight computer-vision pipeline for detecting improperly polished regions on reflective aluminium, developed from a limited dataset and tested on legacy hardware.
At a glance
Context
AI-assisted quality inspection
My focus
Small-data object detection
Tools
Darknet, compact CNNs, CUDA and Python
The challenge
Reflective metal changes appearance with lighting and viewing angle, while specialised industrial datasets can be small or unavailable. The goal was to test whether a compact detector could recognise improperly polished aluminium on modest computing hardware.
My contribution
I led the study from dataset design and software implementation through validation, analysis and publication. The work focused on making defect detection practical under both data and computing constraints.
Approach
- Create and annotate a single-class dataset of deliberately unpolished regions on aluminium samples.
- Use augmentation to explore robustness under limited-data conditions.
- Train a compact Darknet CNN on a workstation and deploy it back to the target low-performance computer.
- Compare CPU and GPU inference on 416 × 416 images.
Prototype evidence
The published experiment used 26 source images and reported 15 FPS with the target GPU, versus 2 FPS on CPU. These figures describe that test setup; the small dataset and single defect class do not establish production-level generalisation.
Open materials and publications
- From Dataset Creation to Defect Detection — Machines, 2024.
- A Comparative Analysis on a Limited Image Dataset — IEEE/ASME MESA, 2024.
- SPADD dataset and model weights — GitHub.