ResidualFormer: A Hybrid CNN-Based Encoder with SegFormer-Inspired Decoder for Efficient Nanoparticle Semantic Segmentation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Nanoparticles are particles with dimensions typically below 100 nanometers, exhibiting unique physical, chemical, and functional properties compared to their bulk counterparts.The segmentation of nanoparticles in Scanning Electron Microscopy (SEM) images is a critical task in material science, as it enables precise characterization of particle morphology, distribution, and surface features.Conventional methods often fail to capture fine-grained particle boundaries and lack robustness against scale and intensity variations, limiting their applicability in large-scale, high-throughput nanoparticle analysis.To address these limitations, this study proposes ResidualFormer, a hybrid deep learning (DL) framework for accurate and efficient nanoparticle segmentation.The model comprises a CNN-based encoder with residual blocks to capture rich spatial representations and a SegFormer-inspired decoder to fuse multi-scale feature maps for precise boundary localization.The dataset comprising TiO₂ SEM images underwent extensive augmentation to improve generalization, and performance was validated using 5-fold cross-validation to ensure robustness.Experimental evaluation demonstrated the superior performance of ResidualFormer, achieving a mean dice score of 0.9540, mean Intersection over Union (IoU) of 0.9121, and mean AUC-ROC of 0.9655, with minimal variability across folds.The standard deviation for dice score was 0.00126, while IoU recorded a standard deviation of 0.002329, reflecting the model's stability across validation splits.These results confirm that the model consistently achieves high accuracy while maintaining robustness against dataset variability.By overcoming the limitations of conventional methods, ResidualFormer establishes a novel, reliable, and efficient solution for nanoparticle segmentation in SEM imagery with the potential to accelerate quantitative material analysis in research and industry.