The right choice depends on the information hidden inside the defect. The goal is not to use the most advanced technology. It is to use the simplest architecture that can make a reliable decision on the real production line.
2D machine vision is often the first choice for presence, orientation, print quality, code reading, contour checks and dimensions visible from a controlled camera view.
3D sensing adds depth information. It is useful where a flat image cannot reliably distinguish height, volume, profile, deformation or surface geometry.
Deep learning can help when acceptable products vary naturally or defects have complex visual patterns. It still requires representative data, controlled acquisition and careful validation.
We start with the acceptance criterion and the failure mode. If a stable threshold, edge, pattern or measurement can separate good from bad parts, a rules-based vision system may be easier to validate and maintain. If the relevant difference is geometric, depth sensing can remove ambiguity that no amount of 2D image processing will solve. If the defect varies in appearance and cannot be described consistently, deep learning may be the right classification layer.
Hybrid systems are common. A station can use conventional vision for dimensions, 3D for geometry and a learned model for a difficult cosmetic region, while storing all results in one traceability record. The architecture should reflect production risk, cycle time, maintenance needs and the evidence required to accept or reject a part.
English English 
Portuguese
French
Spanish