Work

GWOEdgeNet: An ultra-lightweight neural network for efficient real-time Rosa damascena maturity classification on edge devices

Year: 2026
Type: article
Abstract: Accurate and timely maturity classification of Rosa damascena is important for harvest scheduling and for the quality of rose-derived products, but deploying deep models on edge hardware is constrained by compute and memory budgets. We present GWOEdgeNet, a lightweight convolutional network obtained through Grey Wolf Optimization (GWO)-based neural architecture search and trained with Decoupled Knowledge Distillation (DKD) from a MobileNetV2 teacher. The search returned a student with only 10,946 parameters. With an adaptive temperature and distillation weight schedule, DKD raised the 5-fold cross-validation accuracy from 94 . 57 % ± 1 . 19 to 95 . 15 % ± 0 . 96 , and the student reached its peak accuracy 1 . 9 × earlier than the non-distilled baseline. The distilled student retained 97.3% of the teacher’s accuracy at a 203 × compression ratio. On a Raspberry Pi 4B, GWOEdgeNet-KD ran at 38.1 ms per image (26.2 FPS) and was the only model in our comparison that exceeds the 24 FPS real-time threshold, against eight state-of-the-art baselines (MobileNetV3, ShuffleNetV2, EfficientNet-B0, and others) measured under the same protocol. The student was further evaluated under Gaussian noise ( σ ∈ { 0 . 01 , … , 0 . 20 } ) and random occlusion (5–40%). These results indicate that combining metaheuristic NAS with decoupled distillation offers an accurate, compact and real-time option for device agricultural image classification.
Language: English
Open Access status: gold
Estimated APC paid: $1,790
OpenAlex ID: W7166301456
FWCI:
Cites: 20
Cited by:
Related to: 10