Portable Detector Multi-Objek dan Estimasi Susut Bobot Cabai Rawit Pascapanen Berbasis YOLOv8n dan MobileNetV3Small pada Raspberry Pi 5

Erika Dyah Tri Septian, Wijayanti Dwi Astuti, S.Si., M.Sc., Ph.D.

2026 | Tugas Akhir | D4 Teknologi Rekayasa Instrumentasi dan Kontrol

Cabai rawit (Capsicum frutescens) merupakan komoditas hortikultura yang rentan mengalami penurunan mutu pascapanen akibat aktivitas fisiologis seperti respirasi dan transpirasi. Pemantauan susut bobot secara berkala sulit dilakukan dengan metode konvensional karena bersifat merusak sampel. Penelitian ini bertujuan merancang sistem edge vision berbasis deep learning yang mampu mendeteksi multi-objek cabai rawit sekaligus mengestimasi persentase susut bobot secara non-destruktif pada perangkat portable Raspberry Pi 5. Sistem mengombinasikan YOLOv8n untuk deteksi objek dan MobileNetV3Small berbasis tekstur (grayscale dan CLAHE) yang dikonversi ke TensorFlow Lite. Data dikumpulkan dari 200 sampel selama sepuluh hari menghasilkan 6.000 data regresi dan 2.332 gambar dengan 12.064 objek untuk deteksi. Model YOLOv8n mencapai mAP50 0,97, precision 0,93, dan recall 0,92. Model regresi tekstur menghasilkan MAE 6,11%, RMSE 7,91%, dan R² 0,89, lebih baik dibandingkan model warna (MAE 6,30%, RMSE 8,10%, R² 0,88), dikonfirmasi analisis Grad-CAM. Sistem beroperasi real-time dengan waktu pipeline 133,7–167,9 ms dan FPS stabil 22,0 fps. Hasil penelitian menunjukkan bahwa sistem edge vision berbasis fitur tekstur dapat menjadi alternatif non-destruktif yang cepat dan aplikatif untuk pemantauan mutu pascapanen cabai rawit di tingkat petani maupun pedagang pada perangkat dengan sumber daya komputasi terbatas.

Cayenne pepper (Capsicum frutescens) is a horticultural commodity susceptible to postharvest quality deterioration due to physiological activities such as respiration and transpiration. Periodic weight loss monitoring is difficult to perform using conventional methods as they are destructive and prevent further observation of the same sample. This study aims to design an integrated edge vision system based on deep learning capable of performing multi-object detection and non-destructive weight loss percentage estimation of cayenne pepper on a portable Raspberry Pi 5 device. The system combines YOLOv8n for object detection and a texture-based MobileNetV3Small model (grayscale + CLAHE) converted to TensorFlow Lite format. Data were collected from 200 samples over ten days, yielding 6,000 regression data pairs and 2,332 images with 12,064 annotated objects for detection training. The YOLOv8n model achieved mAP50 of 0.97, precision of 0.93, and recall of 0.92. The texture-based regression model outperformed the color-based model with MAE 6.11%, RMSE 7.91%, and R² 0.89 compared to MAE 6.30%, RMSE 8.10%, and R² 0.88, confirmed by Grad-CAM analysis. The system operates in real-time with an average pipeline time of 133.7–167.9 ms and a stable camera FPS of 22.0. The results demonstrate that the texture-based edge vision system can serve as a fast and applicable non-destructive alternative for postharvest quality monitoring of cayenne pepper for farmers and traders on computationally constrained devices. 

Kata Kunci : cabai rawit, susut bobot, deteksi objek, YOLOv8n, MobileNetV3Small, edge vision, non-destruktif, pascapanen