Metallographic Image Segmentation Method Based on Superpixels Algorithm and Transfer Learning

In the production process of aluminium alloy, we can evaluate the microstructure and characterization of alloy metals by using optical microscope and transmission electron microscope. Specifically, we could access the second phase distribution and grain of alloy surface by optical microscope, which are significant indications of alloy qualities and mechanical properties. In the actual industrial production process, the majority of meaningful information are extracted by manual counting. This kind of complicated statistics greatly increase the work of quality inspectors and researchers, which make them hard to concentrate on research of properties of alloy metal. Based on this, we proposed a deep learning method that utilizing U-Net to realize the accurate segmentation of the alloy matrix and the second phase in the metallographic image based on superpixels algorithm and transfer learning. Researchers can easily get some important information about distribution and proportion of the second phase to improve work efficiency. The final experiments indicate that the method proposed in this paper can effectively distinguish alloy matrix and second phase in metallographic image and achieve good results on accuracy, recall, precision and generalization.