AI-Based Surface Defect Detection

In surface inspection, lighting matters more than the model. Scratches on bright sheet metal become visible with low-angle (dark-field) lighting, holes in film and fabric with backlighting, and stains with diffuse front lighting. The camera is triggered by the line encoder: an area-scan camera captures a frame every time the material advances by a set distance, and overlapping frames cover the entire surface. On wide lines, several cameras are mounted side by side.

The model is built in one of two ways. If the defect classes are known and there are enough examples of each, a YOLO-family detection or segmentation model returns both the location and the type of the defect. If defects are rare and examples are scarce, an anomaly detection model that learns only from defect-free surfaces flags every region that deviates from the normal texture. The model runs on the GPU of the edge computer with TensorRT; the design requirement is that processing one frame takes less time than the interval between two camera triggers.

As soon as a defect is found in the image, it is matched to the encoder position. This way, when the defect reaches a marking head or reject point several metres downstream of the camera, it is marked in the right place. For each defect, the metre position, cross-web position, class and an image crop are logged; at the end of each roll or batch, a defect map and quality report are produced.

In concrete paving block and kerb production (vibro-press lines), inspection takes place right after the press, while the product is still wet: as the pallet leaves the press for the curing chamber, the camera looks at the top face of every block and finds broken corners, cracks, incomplete filling, surface voids and defects in the coloured face mix. Because the wet surface is shiny and its tone changes as it dries, lighting geometry and a model trained on your own products are decisive here. Results are logged against the pallet number and shown to the operator in real time; the defect trend reveals a drift in filling, vibration or mix before the whole batch is affected. Product height is measured at the same station with a laser line and a camera.

The camera image reaches the edge AI computer over GigE. The Nexcom COEUS-3721T we hold in stock runs Ubuntu-based JetPack Linux on an NVIDIA Jetson Orin Nano 4 GB module; the model runs on the GPU with TensorRT. The decision goes to the line PLC either directly through the device's digital outputs or via Modbus TCP and OPC UA; the line encoder and the marking unit are handled on the PLC side.

Camera selection depends on line speed. The Crevis MG-A121L-9 in stock (4024 × 3036, colour, 1/1.7 in) has a rolling shutter sensor with a Global Reset mode: it gives good results on stationary or slow-moving surfaces, and on moving surfaces when Global Reset mode is combined with short strobe lighting. For fast-moving material, the global shutter MG-A121K-9 in stock (Sony Pregius IMX304, 4112 × 3008) or a line-scan camera is selected; wide, fast lines need an edge PC with a more powerful GPU.

What decides the success of an AI project is your product, your lighting and your line speed, not a catalogue figure. That is why every project starts with your own samples: first, images are collected on the bench with the right camera, lens and lighting, an initial model is trained, and the false reject and miss rates are measured on your parts. If the result meets the acceptance criterion, the system runs on the line as a pilot, initially in a mode that only logs and makes no decisions.

In surface inspection, during the pilot phase, images captured at the real line speed and under the real lighting are labelled and added to the model; reflections, dust and roll splices that are not seen on the bench come to light at this stage. The defect map is compared with the defects found manually by your quality team; the decision-making mode is enabled once this comparison is complete.

A surface inspection system is usually added to a running line. The camera and lighting are mounted on a bridge, the line encoder signal is shared, and a few signals are added to the existing PLC for marking or rejection. The changes to the line's control software are minor.