Cross-scale attention network for automated carbon nanomaterial recognition in TEM images

Automated interpretation of transmission electron microscopy (TEM) images for nanomaterial classification remains challenging due to complex multi-scale structural patterns, heterogeneous imaging conditions, and limited annotated data. Conventional convolutional neural networks (CNNs) typically operate on single-resolution inputs and single-stream architectures, which restrict their ability to simultaneously capture global morphology and fine structural details essential for distinguishing similar nanomaterial classes. To address these limitations, we propose a multi-resolution, multi-depth CNN architecture with cross-resolution fusion and an auxiliary supervised attention module (ASAM). The framework processes images through parallel resolution-specific branches and integrates features bidirectionally across scales and depths using max-pooling and transposed convolution operations. Auxiliary supervision is introduced at intermediate depths to guide discriminative feature learning and enhance convergence stability. The proposed method was evaluated on a dataset of 5,323 TEM images comprising five categories of carbon-based nanostructures using five-fold cross-validation. Experimental results show that the proposed model achieves an overall accuracy of 94.8%, precision of 94.1%, recall of 94.5%, and F1-score of 94.3%, outperforming baseline models including VGG-16 (84.3% accuracy), ResNet-50 (88.5%), U-Net (90.2%), and EfficientNet (90.3%). Ablation analysis further demonstrates that the auxiliary supervised attention module improves accuracy from 92.5% to 94.8%. These results indicate that the proposed framework provides a robust and scalable expert system for high-throughput TEM-based nanomaterial classification.