A lightweight transfer learning framework for real-time image classification in resource-constrained systems
2026, vol.18 , no.2, pp. 59-70
Article [2026-02-06]
This study investigates an EfficientNetB0-based transfer-learning approach for low-latency image classification on computing platforms with limited resources. The approach is evaluated on a four-class maize leaf disease dataset containing Blight, Common Rust, Gray Leaf Spot, and Healthy samples. The experimental protocol includes stratified dataset partitioning, training-time image augmentation, class-weighted optimization, and a two-stage training strategy in which the pretrained backbone is first frozen and then partially fine-tuned. Model performance is assessed using accuracy, macro-F1, balanced accuracy, ablation experiments, robustness tests, and deployment-related indicators. On the held-out test set, the model achieved 96.34% accuracy and a macro-F1 score of 0.95. Robustness evaluation under brightness changes, contrast shifts, Gaussian noise, and blur showed only limited performance degradation. The final model required approximately 28 MB of storage and achieved an inference time of about 3 ms per image. These results suggest that a compact transfer-learning model can support accurate and computationally efficient image classification in environments where memory, processing capacity, and response time are constrainеed.
real-time image classification, transfer learning, EfficientNetB0, resource-constrained systems, maize leaf disease detection
https://doi.org/10.59035/YLQP1789
Alketa Hyso, Dezdemona Gjylapi. A lightweight transfer learning framework for real-time image classification in resource-constrained systems. International Journal on Information Technologies and Security, vol.18 , no.2, 2026, pp. 59-70. https://doi.org/10.59035/YLQP1789