Featured Application Feature Applications: This paper proposes a data-driven approach based on convolutional neural networks to measure the Vickers hardness value directly from the image of the specimen to get rid of the requirement of the manually generation of indentations for measurement. Hardness testing is an essential test in the metal manufacturing industry, and Vickers hardness is one of the most widely used hardness measurements today. The computer-assisted Vickers hardness test requires manually generating indentations for measurement, but the process is tedious and the measured results may depend on the operator's experience. In light of this, this paper proposes a data-driven approach based on convolutional neural networks to measure the Vickers hardness value directly from the image of the specimen to get rid of the aforementioned limitations. Multi-task learning is introduced in the proposed network to improve the accuracy of Vickers hardness measurement. The metal material used in this paper is medium-carbon chromium-molybdenum alloy steel (SCM 440), which is commonly utilized in automotive industries because of its corrosion resistance, high temperature, and tensile strength. However, the limited samples of SCM 440 and the tedious manual measurement procedure represent the main challenge to collect sufficient data for training and evaluation of the proposed methods. In this regard, this study introduces a new image mixing method to augment the dataset. The experimental results show that the mean absolute error between the Vickers hardness value output by the proposed network architecture can be 10.2 and the value can be further improved to 7.6 if the multi-task learning method is applied. Furthermore, the robustness of the proposed method is confirmed by evaluating the developed models with an additional 59 unseen images provided by specialists for testing, and the experimental results provide evidence to support the reliability and usability of the proposed methods.
Date:
2022-10-22
Relation:
Applied Sciences. 2022 Oct 22;12(21):Article number 10820.
