Abstract
In order to improve the recognition accuracy and model interpretability of metal fracture scanning electron microscope (SEM) images, this research presents an improved dual-branch model (IDBM) based on multi-scale feature fusion. This model employs VGG19 and Inception V3 as parallel branches to separately extract local texture features and global semantic features. Furthermore, it integrates channel and spatial attention mechanisms to enhance the responsiveness of discriminative regions. By integrating dual-branch features using a fixed fusion ratio of 0.8:0.2, the model was trained and validated on an image dataset comprising 800 representative fracture surface images across four categories: cleavage, dimple, fatigue, and intergranular fracture. The results indicate that under small-sample data conditions, the IDBM achieves a Validation Accuracy (Val ACC) of 99.50%, a Recall rate of 99.51%, and an Area Under The Curve (AUC) value of 0.9998, significantly outperforming single models and other fusion strategies. Through integration with class activation mapping (CAM) and feature space visualization analysis, the model exhibits strong interpretability. Furthermore, scale adaptability tests reveal that IDBM maintains stable recognition performance across a magnification range of 100 to 10,000 times, and identifies the optimal observation magnification ranges for the four types of fractures.