Graphite Particle Segmentation Method for Spheroidal Graphite Cast Iron Based on Improved DeepLabv3+

The nodularity in spheroidal graphite cast iron is an important indicator for assessing its quality and performance. It can be utilized for quality control, performance prediction, process optimization, and failure analysis of spheroidal graphite cast iron. Visual analysis of spherical graphite in metallographic samples using optical microscopy has been the most widely adopted method employed by experts to evaluate the nodularity. However, manual evaluation involving the manual counting or measurement of spherical graphite's quantity, size, or other relevant parameters is both time-consuming and costly. While existing rapid methods for calculating the nodularity exist, they often rely on traditional grayscale-based image segmentation algorithms that tend to misclassify similar-colored impurities as graphite particles, resulting in unreliable nodularity calculations. To address this challenge, we propose a deep learning-based approach using DeepLabv3+ (a semantic segmentation network) to extract advanced semantic information from spheroidal graphite cast iron, enabling intelligent segmentation of graphite particles. Additionally, considering the scarcity of metallographic samples of spheroidal graphite cast iron and the insufficient fine-grained edge segmentation capability of semantic segmentation networks for graphite particles, improvements were made to the DeepLabv3+. Finally, our method achieves a 93.50% IoU for graphite particle segmentation on a self-constructed test set, representing a 6.91% improvement compared to the original DeepLabv3+. This research overcomes the challenges in automated evaluation of nodularity in spheroidal graphite cast iron, providing robust support for enhancing the accuracy of quality control and performance prediction in castings.