A high-precision automatic recognition method based on target detection for nanometer scaled precipitates or carbides in different alloys

Abstract

The recognition and statistical analysis of nanometer scaled second phase (NSP), such as precipitates or carbides in alloys is considered to be very important for optimization of mechanical properties. Statistical analysis of the size, number, and distribution of NSPs is required, and manual statistics are difficult and inefficient. In order to solve the above problems, this paper provides an object detection technology for the automatic identification of the NSPs in different alloys. A dataset is constructed by labeling the NSPs in typical transmission electron micrographs (TEM) photographs using a labelImg software. First of all, based on the YOLOv3 object detection network in deep learning, changing the network input size is compatible with the TEM image size. Secondly, using two improved schemes, a priori box K-means clustering and CBAM attention mechanism, a variety of class NSPs detection network is proposed. The statistical results show that the network developed herein has strong applicability. This network can be used for identification for the NSP in several types of alloys, i.e., Cu–Sc alloy, 4Cr5MoSiV1 and Cu–Ni–Si alloy. The accuracy of this network for the recognition of these NSP in Cu–Sc alloy (Cu4Sc precipitates), 4Cr5MoSiV1 (carbides) and Cu–Ni–Si alloy (Ni2Si precipitates) is 93.55%, 81.82% and 93.13%, respectively.

Keywords

• Precipitates
• Object detection
• Deep neural networks
• Cu–Sc alloy
• 4Cr5MoSiV1 steel
• Cu-Ni-Si alloy