A Hybrid CNN-Based Attention Model for Weed Recognition Under Variable Field Conditions with Edge Device Simulations
Journal of Innovative Image Processing · 14.07.2026 · 8 (3) · 938-954
DOI: 10.36548/jiip.2026.3.010 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 14.07.2026. Verfasst von Dharmendra Kumar und 3 weiteren Autor:innen. Die Arbeit nennt 26 Literaturangaben.
Abstract (Originalsprache)
Biotic stress significantly affects agricultural productivity, and weed infestation remains a major factor limiting crop yields and resource-use efficiency. Traditional weed management tools and practices are less effective and labor-intensive at scale. Therefore, automated and intelligent weed identification and classification technologies are required to manage weed infestation. Recent deep learning studies have shown promising results for weed classification. Still, many deep learning architectures have large parameter counts and require high computational power, limiting their deployment in real-world agricultural environments. In this study, a novel lightweight deep learning model, Lightweight Hybrid Channel-Spatial Attention (L-HCSA), is presented for crop-weed image classification in a real-world agricultural environment. L-HCSA is built on a pre-trained MobileNetV2 backbone, enhanced with a hybrid channel-spatial attention module to handle background noise and enhance the discriminative power of weed features. The CWD30 database is used to benchmark the L-HCSA model. An ablation study was performed on five configurations, confirming that the HCSA attention block is the primary contributor to performance. The HCSA attention block adds 0.41 million parameters to the MobileNetV2 baseline (2.63 million parameters) and improves accuracy by 2.18%. The L-HCSA is compared against nine baseline models, including MobileNetV2, MobileNetV3-L, DenseNet-121, ResNet-50, ResNet-101, EfficientNetV2-M, EfficientNet-Lite, TinyViT, and EdgeNeXt. The L-HCSA model has only 3.04 million trainable parameters (12.1 MB) and achieves 77.34% accuracy with a ROC-AUC of 0.9955. A pairwise Wilcoxon signed-rank test with Bonferroni correction and Cohen’s d effect size was performed for statistical evaluation. The L-HCSA model is evaluated for real-time performance, meeting the >=10 FPS threshold in an embedded hardware simulation while maintaining competitive accuracy, achieving 22.0 FPS on the Google Coral Edge TPU.
Zitieren
APA
Kumar, D., Dhankhar, S., Hiranwal, S., & Sisodia, P. S. (2026). A Hybrid CNN-Based Attention Model for Weed Recognition Under Variable Field Conditions with Edge Device Simulations. Journal of Innovative Image Processing, 8(3), 938-954. https://doi.org/10.36548/jiip.2026.3.010
Vancouver
Kumar D, Dhankhar S, Hiranwal S, Sisodia PS. A Hybrid CNN-Based Attention Model for Weed Recognition Under Variable Field Conditions with Edge Device Simulations. Journal of Innovative Image Processing. 2026;8(3):938-954. doi:10.36548/jiip.2026.3.010
BibTeX
@article{Kumar2026,
author = {Kumar, Dharmendra and Dhankhar, Sunil and Hiranwal, Saroj and Sisodia, Pushpendra Singh},
title = {A Hybrid CNN-Based Attention Model for Weed Recognition Under Variable Field Conditions with Edge Device Simulations},
journal = {Journal of Innovative Image Processing},
year = {2026},
volume = {8},
number = {3},
pages = {938--954},
doi = {10.36548/jiip.2026.3.010}
}
EUVIMED