Automated Vickers hardness measurement using convolutional neural networks

Abstract

A method based on convolutional neural networks (CNN) was proposed for robust automated measurement of Vickers hardness. Vickers hardness testing is generally applied to metals but can be conducted on other materials such as ceramics. Typical image processing methods tend to fail to detect the indentations in materials with rough surfaces, noisy patterned surfaces, distorted indentation shapes, or cracks. Automated measurement methods must be robust and versatile to process materials with these anomalies. A CNN method was chosen to detect the bounding box of the Vickers indentation, as the technique automatically extracts features required to recognize the region of an object. The diagonal lengths and Vickers hardness values calculated using the CNN method were similar to the reference manual measurements, even in samples with rough surfaces, noisy patterned surfaces, distorted indentation shapes, or cracks. The CNN method is a promising automated alternative for robust and versatile prediction of Vickers hardness.