Avoiding Post-Processing with Context: Texture Boundary Detection in Metallography

Accurately identifying grain boundaries in metallographic images is challenging due to the intricate nature of texture boundaries. State-of-the-art (SOTA) models, like the Segment Anything Model (SAM), often fail in purely texture-based segmentation tasks without clear object boundaries. The specific case of models like SAM also requires prompts which in this context requires prior knowledge of grain position so the model can be seeded. Moreover, manual annotation is not only time-consuming but also subjective and context-sensitive. Current SOTA methods rely on small annotated patches for training and require extensive post-processing during inference to merge patch boundary maps. This approach often leads to overfitting to the ground truth and results in models that are not well-generalized. We introduce MLOgraphy++, a novel approach that eliminates the need for post-processing by training on partially labeled context windows. Our method leverages a U-Net architecture trained with large context windows, where only a small portion is annotated, allowing the model to learn boundary segmentation in context. During inference, our model effectively handles partial and incomplete boundaries while accommodating context variations without the need for post-processing. To evaluate our approach, we adopt the Heyn intercept method, a classical technique for measuring average grain size, as a more suitable metric than pixel accuracy, and apply it to MLOgraphy++ and a fine-tuned AutoSAM model. This method better captures the critical distribution of grain sizes, which is difficult to label accurately on a pixel level. We benchmark MLOgraphy++ against the SOTA MLOgraphy on the Texture Boundary in Metallography (TBM) dataset. Our results demonstrate that MLOgraphy++ achieves comparable performance while eliminating the need for post-processing, thus enhancing the generalizability of the method. This work highlights the importance of contextual training in improving the accuracy and practicality of texture boundary detection in metallography.Source code and dataset: https://github.com/Scientific-Computing-Lab/MLOgraphyPlusPlus