Accurate identification of preparation defects in 20 steel (equivalent to AISI 1020) metallographic specimens is the key to ensuring the reliability of microstructural rating. However, under high-magnification microscopic fields of view, weak defect features often suffer from severe aliasing with high-frequency, strongly structured grain boundary texture backgrounds. This causes conventional detectors to face serious issues of feature dilution and over-smoothing during feature extraction. To address this challenge, this paper proposes a lightweight recognition framework fusing structured priors and energy-aware attention, named RSS-YOLOv10. This method focuses on feature decoupling within local high-resolution fields of view. First, a RepHMA module is designed to construct a heterogeneous multi-scale receptive field, capturing defect morphologies with extreme scale variations through a structural re-parameterization mechanism. Second, a parameter-free SimAM attention mechanism based on energy functions is introduced. Leveraging the principle of information entropy, it adaptively suppresses high-frequency grain boundary noise in the feature space, significantly enhancing the signal-to-noise ratio (SNR) of weak defects against strong texture backgrounds. Finally, the feature aggregation path is reconstructed using SlimNeck to achieve a lightweight architecture. Experimental results demonstrate that on datasets retaining original micro-texture details, RSS-YOLOv10 exhibits superior identification capabilities, achieving mAP scores of 80.1% and 78.9% on a self-constructed dataset and the NEU-DET dataset, respectively. Compared with baseline models, this method effectively overcomes the bottleneck of missed detections under complex grain boundary backgrounds while reducing computational complexity (GFLOPs) and parameter redundancy, providing an efficient solution for the fine screening of challenging defect samples in metallographic analysis.
Shen et al. (Thu,) studied this question.