Electronic Theses and Dissertation

Klasifikasi mutu biji kopi secara manual memerlukan waktu relatif lama dan rentan menghasilkan penilaian yang tidak konsisten, terutama dalam pemisahan biji kopi normal dan cacat. Penelitian ini bertujuan membandingkan arsitektur deep learning untuk klasifikasi biner biji kopi berbasis multi-dataset serta menentukan konfigurasi model terbaik pada lingkungan cloud-based dan Edge AI. Dataset yang digunakan adalah USK-Coffee Dataset sebanyak 8.000 citra dan Coffee Green Bean with 17 Defects Dataset sebanyak 979 citra, yang labelnya diharmonisasi menjadi kelas normal dan cacat. Metode transfer learning diterapkan pada MobileNetV3-Large, MobileViTv2-100, EfficientNetV2-S, dan FocalNet-Tiny-SRF. Model dikonversi ke format ONNX FP32, FP16, INT8 Static, dan INT8 Dynamic, kemudian dievaluasi pada cloud-based CPU, cloud-based CUDA, edge AI-based CPU, edge AI-based CUDA, dan edge AI-based TensorRT. Evaluasi mencakup akurasi, precision, recall, F1-score, specificity, ROC-AUC, throughput, latensi P90, ukuran model, konsumsi energi, emisi karbon, SCI, serta pemeringkatan CRITIC dan TOPSIS-MMD. Hasil menunjukkan bahwa seluruh model FP32 memperoleh akurasi di atas 96% dan ROC-AUC di atas 0,98. Konfigurasi terbaik adalah MobileNetV3-Large FP16 pada Jetson Orin Nano 4GB dengan TensorRT, dengan akurasi 97,48%, ROC-AUC 0,9862, latensi P90 11,03 ms/gambar, ukuran model 8,05 MiB, dan SCI 6,89×10⁻⁶ ± 3,72 × 10⁻⁷ g CO₂e/citra. Hasil ini menunjukkan bahwa konfigurasi tersebut memberikan keseimbangan paling baik antara ketepatan klasifikasi, efisiensi inferensi, ukuran model, intensitas karbon, dan kesiapan deployment pada perangkat edge AI.

Kata kunci: klasifikasi biji kopi, deep learning, Edge AI, ONNX dan kuantisasi model, Software Carbon Intensity, TOPSIS-MMD

Manual coffee bean quality classification is relatively time-consuming and prone to inconsistent assessment, particularly in separating normal and defective coffee beans. This study aims to compare deep learning architectures for binary coffee bean classification using a multi-dataset approach and to determine the best model configuration for cloud-based and Edge AI environments. The datasets used were the USK-Coffee Dataset, consisting of 8,000 images, and the Coffee Green Bean with 17 Defects Dataset, consisting of 979 images, whose labels were harmonized into normal and defective classes. A transfer learning method was applied to MobileNetV3-Large, MobileViTv2-100, EfficientNetV2-S, and FocalNet-Tiny-SRF. The models were converted into ONNX FP32, FP16, INT8 Static, and INT8 Dynamic formats, then evaluated on cloud-based CPU, cloud-based CUDA, edge AI-based CPU, edge AI-based CUDA, and edge AI-based TensorRT. The evaluation covered accuracy, precision, recall, F1-score, specificity, ROC-AUC, throughput, P90 latency, model size, energy consumption, carbon emissions, SCI, and ranking using CRITIC and TOPSIS-MMD. The results show that all FP32 models achieved accuracy above 96% and ROC-AUC above 0.98. The best configuration was MobileNetV3-Large FP16 on Jetson Orin Nano 4GB with TensorRT, achieving 97.48% accuracy, 0.9862 ROC-AUC, 11.03 ms/image P90 latency, 8.05 MiB model size, and 6.89×10⁻⁶ ± 3,72 × 10⁻⁷ g CO₂e/image SCI. These results indicate that this configuration provides the best balance among classification accuracy, inference efficiency, model size, carbon intensity, and deployment readiness on Edge AI devices. Keywords: coffee bean classification, deep learning, Edge AI, ONNX model quantization, Software Carbon Intensity, TOPSIS-MMD