Cognex Deep Learning defect detection tool reliably detects pores in metal
A breakthrough in complex inspection, part location, classification, and OCR
The cylinder block is the foundation of an automotive engine. Its large cylinders are the primary working parts of a reciprocating engine, built to hold the pistons as they pump up and down under compression. Cylinders are normally made of cast metal and are sometimes lined, or “sleeved,” in a lubricating coating. The cylinder walls must be durable as they make contact with a piston’s compression rings. The cylinder can tolerate small irregularities in its metals, which are smoothed out during mechanical wear, but must not contain bubbles in the metal known as “pores.” These defects are especially challenging to inspect because a cylinder’s surface is rough, and images appear blurry around the edges due to the depth of field. Specular lighting, or glare, on the cylinder’s reflective surface also complicates the inspection. It is difficult to program an automated inspection that tolerates so many small variations in feature shape and location, as well as glare and blurriness.
Cognex Deep Learning quickly identifies pores when other methods struggle to inspect under the same lighting conditions. Within minutes, an engineer can train the software on a representative set of “good” and “bad” images of a cylinder, adjusting the region of interest with a masking filter to eliminate the bright disk of negative space in the shaft. Using the defect detection tool in supervised mode, a technician annotates the pores in the images labeled as “bad” and adjusts parameters, including feature size, scale, aspect ratio, and shear to help the model account for variations in appearance. “Good” images which depict normal cylinders help the software learn which types of minor casting anomalies and variations are tolerable. The engineer can retrain the system, adjusting parameters and adding additional images, until the model can generalize the normal appearance of a cylinder and recognize the abnormalities. During run-time, the deep learning-based software inspects each image within milliseconds, characterizing those with pores as defective and the rest as normal.