Vision system for cast and machined parts: a multi-camera inspection station that finds casting surface defects on the line

01

Eight casting surface defects and the depth at which each counts

A 2025 Scientific Data paper describing the Metal Surface Defect Dataset (MSDD) of aluminium alloy casting defects separates real defects from harmless surface features by depth: a low-reflectivity stain is not an inclusion, but a hole deeper than 0.5 mm is; tool-vibration lines are not cut marks, but marks deeper than 0.1 mm are. [1]

Give the supplier sample parts with each defect class you reject and good parts with the harmless features that must pass, the depth or size limits for each, the surfaces to inspect, the line rate and part presentation, and how rejects are to be marked or diverted.

02

Strobed lights from all around and cameras at different heights, inside a 300-parts-per-hour cycle

The MSDD acquisition method places multiple CCD cameras and light sources around the part at different heights, and a PLC strobes the lights in sequence with the cameras so each part is imaged under light from different angles; each camera runs at 30 frames per second and the whole capture takes 0.96 s, which the authors state meets the 300 parts per hour of an aluminium alloy casting line. Images from the different light angles are blended into pseudo-colour images for standard object detectors; the dataset holds 138,585 single-channel and 9,239 blended images, four detectors reached 86.1% mAP on it, and the authors report that the method struggles on highly reflective surfaces such as polished metal, where the lights produce reflections. [1]

Editorial calculation: 300 parts per hour leaves 12 s per part, so 0.96 s of image capture uses 8% of the cycle; at 30 frames per second each camera takes about 29 frames in that time. [1]

Hanchine lists bearing surface inspection machines, models HZ-VI-BB0515, HZ-VI-BB1530 and a 3060 model, integrating 2D and 3D vision, alongside bearing roller and ring inspection machines. [3]

03

Detector accuracy and speed on a casting surface defect dataset

A 2024 Sensors study of casting surface defect detection reports [email protected] on its casting dataset of 0.783 for RetinaNet, 0.827 for YOLOv5n, 0.832 for YOLOv8n, 0.835 for Faster R-CNN, 0.846 for Mask R-CNN, 0.854 for TOOD and 0.862 for its own SLGA-YOLO; the YOLO models need 2.5–3.0 million parameters and 7.1–10.9 GFLOPs, against 32–48 million parameters and 199–274 GFLOPs for Mask R-CNN, Faster R-CNN, TOOD and DINO. In its ablation, the baseline YOLOv8n ran at 105 frames per second and SLGA-YOLO at 133. [2]

Editorial note: an mAP near 0.86 means some defects are still missed or mislocated; ask the supplier for detection and false-reject rates per defect class, measured on your own parts across multiple shifts, rather than a dataset mAP.

04

Send the parts, defect limits and line timing together

Paste it into your enquiry below or into an email to a maker, then replace every bracket.

Parts: [alloy; process; surfaces to inspect; finish]
Defects: [classes; depth or size limits; samples]
Line: [parts per hour; presentation; fixturing]
Imaging: [2D; multi-angle strobe; 3D]
Output: [reject signal; marking; data logging]
Acceptance: [detection rate; false-reject rate; trial parts]

See manufacturer records Send an enquiry through CNFX