EUVIMEDEUVIMED entdecken
Crossref

Multimodal Late Fusion Using CNN Backbones of Varying Complexity for Automated Peanut Kernel Size Assessment

Haydar Tuna, A. Burak Güher

Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji · 20.11.2026 · (Advanced Online Publication)

DOI: 10.29109/gujsc.2015299 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji, online erschienen am 20.11.2026. Verfasst von Haydar Tuna und 1 weiteren Autor:innen. Die Arbeit nennt 70 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Peanut kernels are graded by size, which determines quality and market value. However, most vision-based grading systems rely on RGB images alone and overlook the texture and edge details separating size classes. Late fusion of RGB, texture and edge maps can provide these details. Its performance depends on the fusion strategy and CNN architecture, and their combination has received limited attention in peanut grading. To address this gap, this study proposes a decision level fusion framework combining RGB images, Local Binary Pattern (LBP) texture maps and Canny edge maps. Five ImageNet-pretrained CNNs (MobileNetV2, ResNet-18, GoogLeNet, VGG-16 and DenseNet-121) were trained on each stream using a balanced dataset of 2,100 im ages from three size classes. Their outputs were then combined with four fusion strategies under five-fold cross-validation. In all architectures, three-stream fusion outperformed single-stream models, and VGG-16 with simple average fusion achieved the highest accuracy (96.76% ± 0.46%). Background masking further improved the LBP stream by up to 10.67 percentage points. Lightweight MobileNetV2 and ResNet-18 also remained competitive with fewer parameters. Simple average fusion performed best or equal-best in four architectures, making it a practical default. Overall, multimodal late fusion is effective for peanut grading and shows potential for other fine-grained agricultural tasks.

Zitieren

APA

Tuna, H., & Güher, A. B. (2026). Multimodal Late Fusion Using CNN Backbones of Varying Complexity for Automated Peanut Kernel Size Assessment. Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji. https://doi.org/10.29109/gujsc.2015299

Vancouver

Tuna H, Güher AB. Multimodal Late Fusion Using CNN Backbones of Varying Complexity for Automated Peanut Kernel Size Assessment. Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji. 2026. doi:10.29109/gujsc.2015299

BibTeX
@article{Tuna2026,
  author = {Tuna, Haydar and Güher, A. Burak},
  title = {Multimodal Late Fusion Using CNN Backbones of Varying Complexity for Automated Peanut Kernel Size Assessment},
  journal = {Gazi Üniversitesi Fen Bilimleri Dergisi Part C: Tasarım ve Teknoloji},
  year = {2026},
  number = {Advanced Online Publication},
  doi = {10.29109/gujsc.2015299}
}
Kostenlos starten →