Measuring Brinell hardness indentation by using a convolutional neural network

For robust automatic measurement of Brinell hardness, we propose a novel measurement of the indentation diameter by using a convolutional neural network (CNN), which is a type of deep learning. In the Brinell hardness measurement, it is sometimes difficult to detect the indentation edges by image processing methods, due to the surface conditions or a change in the contrast with hardness levels. CNNs can automatically extract features required for object recognition of humans, hence we expect that the CNN can detect the indentation edges as robust as human operators, regardless of the surface conditions. We developed a CNN system to detect the indentation edges, and trained the CNN using the dataset combined with the indentation edge images and the position of edges given by a human operator. To verify the usefulness of the CNN method, we compared this method with measurement by a human operator and the region growing (RG) method, which is a simple region segmentation method in image processing. In the CNN method, indentation diameters and Brinell hardness were in good agreement with the manual measurement compared with the RG method, regardless of the hardness levels of test samples. Moreover, the CNN method enabled good measurement of a rough surface that is different from the dataset for training the CNN. Thus, our novel method is as robust as measurement by experienced operators and versatile in terms of independence of the hardness of samples or the characteristics of the surface conditions.