<![CDATA[Guava cultivation is often threatened by leaf diseases that disrupted plant growth and reduced agricultural productivity. Early and accurate disease identification was crucial but relied heavily on slow and subjective visual inspections by human experts. This study proposed an automated, accurate, and efficient solution by comparing lightweight deep learning models. A total of 656 augmented guava leaf images representing four classes (Algal Leaf Spot, Insects Eaten, Red Rust, and Healthy Leaf) were evaluated. Using a transfer learning approach, the hyperparameters were systematically tuned for MobileNetV3-Small as the proposed model and compared with MobileNetV2 as the baseline architecture. The experimental results demonstrated that MobileNetV3-Small achieved a superior test accuracy of 91.00%, outperforming MobileNetV2, which achieved 87.00%. The integration of Squeeze-and-Excitation (SE) modules and the h-swish activation function in MobileNetV3-Small significantly improved the identification of subtle visual symptoms, particularly for the Healthy Leaf and Insects Eaten classes. However, MobileNetV2 maintained a slight advantage in real-time processing speed (54.86 FPS versus 50.30 FPS) because of memory-bound bottlenecks associated with the SE modules. Overall, MobileNetV3-Small provided superior diagnostic accuracy, whereas MobileNetV2 remained a highly viable option for latency-critical deployment on low-end devices.]]>
Copyrights © 2026
