EE Seminar: Texture Aware Segmentation in Metallography
https://tau-ac-il.zoom.us/j/85097922575?pwd=5k77VVbGUlvcv2IisxJBoLMcTNjEZW.1
Electrical Engineering Systems ZOOM Seminar
Speaker: Inbal Cohen
M.Sc. student under the supervision of Prof. Shai Avidan and Dr. Gal Oren
Tuesday, 9th September 2025, at 15:00
Texture Aware Segmentation in Metallography
Abstract
Accurately identifying grain boundaries in metallographic images is challenging due to the subtle and complex nature of texture-defined transitions. Traditional methods rely on dense, fully annotated data or assume the presence of clear object boundaries—assumptions that rarely hold in texture-driven domains like metallography.We propose two complementary solutions that leverage partial supervision and texture-aware learning.
First, MLOgraphy++ trains a U-Net on large context windows using only partially annotated regions, enabling it to learn boundary cues in realistic settings while reducing annotation effort. It eliminates the need for patch-based inference or post-processing, and we evaluate it using the Heyn intercept method—a more application-relevant metric for grain size distribution.
Second, TextureSAM addresses the limitations of prompt-based models like SAM, which are biased toward object shape. We fine-tune SAM on a texture-augmented ADE20K dataset, guiding it to prioritize texture-defined boundaries. TextureSAM significantly outperforms SAM in both synthetic and real-world texture segmentation tasks.
Together, these methods offer a robust framework for texture boundary detection, advancing scalable and accurate segmentation in texture-critical domains.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום בצ'ט של שם מלא + מספר ת.ז.
