<![CDATA[The ripeness level of bananas is an important indicator that determines the quality, selling value, and suitability of distribution in the agricultural supply chain. However, manual maturity assessments are still subjective and difficult to apply consistently on a large scale. The use of Convolutional Neural Networks (CNN) offers a more accurate and objective solution, but most previous studies have only evaluated high-powered devices so they do not reflect the real performance of low-spec smartphones. This study aims to compare the efficiency of three lightweight CNN architectures: MobileNetV1, EfficientNetB0, and NASNetMobile for the classification of banana ripeness and evaluate its feasibility of being implemented on low-end Android devices. The research method included model training using the Banana Ripeness Classification dataset containing 13,478 images with three maturity classes. Augmentation-based oversampling was applied to address data imbalances, while all three models were trained on transfer learning strategies before being converted to the TensorFlow Lite format. Direct testing was conducted on the Samsung Galaxy A3 (2016) device to measure accuracy, inference time, model size, and RAM usage. The experimental results showed that MobileNetV1 provided the best performance with an accuracy of 98.14%, an inference time of 287.57 ms, and a model size of 3.23 MB, much more efficient than EfficientNetB0 and NASNetMobile. In conclusion, MobileNetV1 is the most optimal architecture for Android-based banana ripeness classification applications on low-spec devices, while making an empirical contribution to the selection of efficient CNN models for mobile implementation in the context of digital agriculture.]]>
Copyrights © 2026
