2025/06/06 by Ashif Ahmed Shuvo, Wahada Jinnat Oishy Bhuian, Afzal Rahman +1 · 1 voice
Computer Science · #IoT and Edge/Fog Computing #Mobile and Web Applications #Advanced Malware Detection Techniques
paper · doi:10.1016/j.atech.2025.101072
openalex publication_date 2025/06/06 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/01
The integration of on-device machine learning (ML) into mobile platforms has the potential to enable intelligent, real-time diagnostics in agricultural settings. This study presents EinsteinNet, a lightweight convolutional neural network (CNN) optimized for offline orange quality classification on Android devices. A custom dataset of 15,000 annotated images across five quality categories—fresh, rotten, green, canker-affected, and black-spotted—was used to train and compare EinsteinNet against four established architectures (ResNet50, DenseNet121, MobileNetV2, NASNetMobile) and a no-code Google Teachable Machine baseline. EinsteinNet achieved 99.6% test accuracy with a compact model size (254 KB), but incurred higher inference latency (∼1118 ms) relative to other models. All networks were converted to TensorFlow Lite (TFLite) format and integrated into an Android application with full offline inference capabilities. Empirical evaluation on a Google Pixel 6 showed that while custom CNNs offer strong classification performance and deployment efficiency, optimizing for real-time responsiveness remains critical. Power consumption metrics collected via Android Profiler revealed critical trade-offs among inference accuracy, latency, and energy usage, underscoring the balance required in deploying edge AI models for precision agriculture.