Analisis Komparatif Arsitektur CNN Untuk Klasifikasi Kelayakan Buah Mangga: ResNet50 Dan MobileNetV2

Abstract

Post-harvest losses of mangoes in Indonesia reach 30–40% due to subjective manual sorting procedures that are unable to detect early-stage pathological anomalies. This study aims to compare the performance of the ResNet50 and MobileNetV2 Convolutional Neural Network (CNN) architectures for mango fruit fitness classification based on the Indonesian National Standard (SNI) 3164:2009. The method used is transfer learning with a dataset of 1,700 images from Kaggle containing cluttered backgrounds to replicate the real conditions of Indonesian agricultural MSMEs, divided into 80% training, 10.6% validation, and 9.4% testing sets. ResNet50 was evaluated in two scenarios (base and fine-tuned), while MobileNetV2 was evaluated without fine-tuning. The results show that MobileNetV2 achieved the highest accuracy of 97% (F1-Score = 0.97) with an inference latency of 7.60 ms and a training duration of ~55 seconds. ResNet50 base only achieved 78% with a latency of 12.70 ms, and ResNet50 fine-tuned achieved 94% with a latency of 12.80 ms and a total two-phase training duration of ~272 seconds. In conclusion, MobileNetV2 is superior to ResNet50 in both accuracy and computational efficiency on cluttered background datasets, making it an optimal candidate to be converted to TensorFlow Lite and implemented in real-time on Indonesian farmers' and MSMEs' smartphones to eliminate post-harvest losses.