<![CDATA[National food security relies heavily on rapid and accurate control of rice plant diseases. However, the implementation of automatic detection technology at the farmer level is often hampered by low image quality due to the use of low-spec mobile phone cameras and data compression in areas with poor signal. This study aims to evaluate the performance of two lightweight Deep Learning architectures, MobileNetV3-Small and EfficientNet-Lite, in classifying rice diseases under low-resolution image conditions. The research method applies a simulation of resolution degradation to 128x128 pixels on a dataset consisting of four classes: Blast, Blight, Brown Spot, and Healthy. Empirical test results show that EfficientNet-Lite is significantly superior in diagnostic accuracy with an accuracy of 95.67%, a precision of 95.65%, and a recall of 95.33%. In contrast, MobileNetV3-Small achieved an accuracy of 89.00%, yet offered superior computational efficiency: a model size of only 11.97 MB (73% smaller than EfficientNet-Lite's 45.48 MB) and an inference speed of 4.67 milliseconds per image, equivalent to 214 frames per second (FPS). The study concluded that EfficientNet-Lite is recommended for high-precision diagnostic systems, while MobileNetV3-Small is the most adaptive solution for real-time applications on storage-constrained mobile devices.]]>
Copyrights © 2026
