Akademik Veri Yönetim Sistemi
Pattern Analysis and Applications, cilt.29, sa.3, 2026 (SCI-Expanded, Scopus)
Machine learning techniques have demonstrated effectiveness across various domains, including agriculture, where classification models have been developed for crop counting, plant recognition, and disease detection. However, computational cost is often overlooked in the design of these models, which may limit the practicality of such models on edge devices. Efficient models with fewer parameters and FLOPs are important for practical use, as they reduce computational and memory demands and may support cost-effective operation under resource constraints. In this study, the Ghost-DeepResNet model was proposed for detecting tipburn, old leaves, and healthy leaves in strawberries. The model incorporates Ghost modules within DeepResNet blocks, resulting in an efficient and accurate architecture. With only 14,479 parameters and 53 MFLOPs, the model achieves a mean accuracy of 96.62% using stratified k-fold cross-validation, which serves as the primary indicator of robustness and within-dataset generalization. In addition, a fixed train–validation–test split produced 99.74% accuracy, serving as the baseline for pruning and quantization. Pruning and quantization yielded a model with 9,495 parameters, 33 MFLOPs, and 0.907 ms mean latency on an NVIDIA GeForce RTX 3060, with a held-out accuracy drop to 98.21%, suggesting that this ultra-lightweight model can still benefit from pruning and quantization. On the Jetson AGX Orin platform, the original and pruned variants achieved FP16 mean latencies of approximately 1.15 ms, indicating potential for low-latency edge inference, while further validation across diverse field conditions and lower-power edge devices remains necessary. Code is available at: https://github.com/MuhabHariri/Ghost-DeepResNet.