SpiceVision: A Dual-Attention EfficientNetB3 Framework with Multi-Scale CBAM Feature Fusion for Fine-Grained Spice Image Classification
International Journal of IoT, Embedded Systems and Industrial Automation · 2026
DOI: 10.66261/nz54mw97 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of IoT, Embedded Systems and Industrial Automation, 2026. Verfasst von Deep Kamle und 2 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Visual classification of raw spices is essential for ensuring food quality and safety, automated sorting in retail settings, culinary education and detection of economic adulteration, but is difficult due to their high similarities in terms of color, granularity, and texture attributes. This work introduces SpiceVision, a light-weight hybrid deep learning framework integrates a pretrained ImageNet EfficientnetB3 backbone network (consisting of its first six MBConv blocks), two Convolutional Black Attention Modules (CBAM) added to two intermediate stages and merging them via an upsample-fusion multi-scale fusion module followed by a classification head using dual global pooling. The input images are preprocessed using a four-step procedure comprising resizing to 128×128. Contrast-Limited Adaptive Histogram Equalisation (CLAHE), and per-channel normalization learned by the model itself. The network was trained with a carefully selected data set consisting of 17,452 images for 20 different spices (12,215 training / 2,619 validation / 2,618 test images, data split 70% / 15% / 15%) via a two-step transfer-learning process of 20-epoches warn-up and cosine decay followed by 30-epoches fine-tuning with low learning rate where only the attention, fusion and classification layers were updated (765,032 out of 5,943,033 parameters, or 12.9%) while keeping EfficientNetB3 completely frozen. With the trained model, 98.62% test accuracy, macro F1-score 98.70%, weighted F1-score 98.62%, and macro one-vs-rest ROC-AUC 0.9999 is achieved. In per-class evaluation, each of the 20 different spices gets a precision, recall, and F1-score greater than 0.94, and the remaining small amount of confusion is almost all within several pairs of texturally similar seed and powder types like cloves vs. cumin seed and coriander vs. coriander seeds.
Zitieren
APA
Kamle, D., Sharma, D. S., & Shrivastava, D. H. (2026). SpiceVision: A Dual-Attention EfficientNetB3 Framework with Multi-Scale CBAM Feature Fusion for Fine-Grained Spice Image Classification. International Journal of IoT, Embedded Systems and Industrial Automation. https://doi.org/10.66261/nz54mw97
Vancouver
Kamle D, Sharma DS, Shrivastava DH. SpiceVision: A Dual-Attention EfficientNetB3 Framework with Multi-Scale CBAM Feature Fusion for Fine-Grained Spice Image Classification. International Journal of IoT, Embedded Systems and Industrial Automation. 2026. doi:10.66261/nz54mw97
BibTeX
@article{Kamle2026,
author = {Kamle, Deep and Sharma, Dr. Satyendra and Shrivastava, Dr. Hemang},
title = {SpiceVision: A Dual-Attention EfficientNetB3 Framework with Multi-Scale CBAM Feature Fusion for Fine-Grained Spice Image Classification},
journal = {International Journal of IoT, Embedded Systems and Industrial Automation},
year = {2026},
doi = {10.66261/nz54mw97}
}
EUVIMED