Vision for sorting systems

Industrial vision joins optics, data, inference, and actuation. A classifier cannot compensate for unstable lighting, motion blur, or a control loop that receives a correct decision after the object has already passed the actuator.

A useful first experiment is deliberately simple: fix the camera, control the light, collect examples from the actual line, and measure the complete time from exposure to action. This reveals whether the real problem is classification, image quality, mechanics, or timing.

Dataset design should follow the failure modes of the environment. Near-duplicate images can make validation look excellent while hiding poor generalisation. I prefer separating data by capture session and testing edge cases such as partial objects, reflections, occlusion, and unexpected backgrounds.

The final metric is not model accuracy alone. It is the quality of the sorting decision over time: correct routing, missed objects, false rejects, throughput, and how quickly an operator can understand a fault.