MET-SE: A Channel Attention Module for High-Fidelity Metallographic Image Generation

This paper presents a metallographic image generation framework designed for microstructure synthesis in the steel industry. To enhance generation quality and efficiency, a channel attention module, Metallography Squeeze-and-Excitation (MET-SE), is embedded into the intermediate layers of a pix2pixHD-based architecture. This module adopts the channel attention mechanism to improve the inter-channel dependency relationship in feature extraction, thereby contributing to more accurate and true-preserving image generation. The proposed approach also marks an effort to adapt generative models specifically to metallographic data, addressing challenges such as complex phase structures and fine grain boundaries. Experiments show that the MET-SE module improves image fidelity while reducing training time and resource consumption. This method offers a stable and efficient solution for generating high-quality metallographic datasets in materials research.