Transformer-Based Hybrid Convolution Attention Framework for Plant Species Classification
Proceedings of Engineering and Technology Innovation · 22.07.2026 · 34 · 47-62
DOI: 10.46604/peti.2026.16186 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of Engineering and Technology Innovation, online erschienen am 22.07.2026. Verfasst von Karnan Arunasalam und 1 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
Fine-grained plant species identification remains challenging because of high intra-class variation and strong inter-class similarity among closely related taxa. Existing hybrid convolutional neural network (CNN)-Transformer architectures progressively attenuate discriminative texture cues during deep attention processing. To address this limitation, this study aims to develop the Transformer-based hybrid convolution attention framework (TB-HCAF), a targeted extension of CoAtNet that preserves complementary local morphological and global structural representations through a dual-path architecture, a parameter-efficient spatial alignment block, and a channel-wise hybrid fusion. The framework is evaluated using a leakage-safe, group-stratified k-fold protocol eliminates specimen-level data leakage and validated through comparative experiments, ablation studies, effective receptive field (ERF) analysis, and cross-dataset evaluation. TB-HCAF achieves high Top-1 accuracy on PlantCLEF 2015, surpassing representative baseline models, and further demonstrates strong zero-shot cross-dataset generalization on Oxford 102 Flowers and iNaturalist 2018. ERF analysis confirms improved discrimination of morphologically similar species and robustness across long-tailed distributions.
Zitieren
APA
Arunasalam, K., & Rengaswamy, R. (2026). Transformer-Based Hybrid Convolution Attention Framework for Plant Species Classification. Proceedings of Engineering and Technology Innovation, 34, 47-62. https://doi.org/10.46604/peti.2026.16186
Vancouver
Arunasalam K, Rengaswamy R. Transformer-Based Hybrid Convolution Attention Framework for Plant Species Classification. Proceedings of Engineering and Technology Innovation. 2026;34:47-62. doi:10.46604/peti.2026.16186
BibTeX
@article{Arunasalam2026,
author = {Arunasalam, Karnan and Rengaswamy, Ragupathy},
title = {Transformer-Based Hybrid Convolution Attention Framework for Plant Species Classification},
journal = {Proceedings of Engineering and Technology Innovation},
year = {2026},
volume = {34},
pages = {47--62},
doi = {10.46604/peti.2026.16186}
}
EUVIMED