Semi-Supervised Learning Framework for Aluminum Alloy Metallographic Image Segmentation

Abstract

The purpose of aluminum alloy metallographic image segmentation is to automatically recognize and segment the microstructures of aluminum alloy, which is an important topic in the fields of materials science research and product assessment. In order to achieve satisfactory segmentation results, we always need to label each pixel in metallographic image. This labeling work is very costly in terms of time and human effort. In this paper, we propose a semi-supervised learning framework which is able to use a small number of labeled images to achieve outstanding performance. Based on this framework, we first implement a semi-supervised segmentation method for aluminum alloy metallographic images. The U-Net is used as the basic segmentation model in this method. Second, we propose a self-paced semi-supervised segmentation method to reduce the interfere caused by the mislabeled images. To further improve the segmentation performance, we embed the improved pseudo-label selection strategies and loss functions in our framework, and implement four improved semi-supervised segmentation methods. In addition, we construct a new aluminum alloy metallographic image dataset to evaluate the proposed semi-supervised learning framework. The experiments indicate that our method is able to achieve very competitive results with a small number of labeled images compared to the state-of-the-art supervised method using a large number of pixel-level labels.