Electronic Theses and Dissertation

Buah-buahan tropis merupakan komoditas hortikultura utama di indonesia yang memerlukan penanganan panen dan pascapanen secara tepat untuk menjaga kualitas, kematangan, dan kesegarannya. proses pengendalian mutu buah umumnya dilakukan secara manual karena keterbatasan sumber daya komputasi, sehingga rentan terhadap kesalahan terutama pada kondisi lapangan yang beragam. penelitian ini dilakukan untuk mengembangkan model multi-task learning (mtl) berbasis pendekatan hybrid yolov12 dan convolutional neural network (cnn) untuk mendeteksi serta mengklasifikasikan jenis buah, tingkat kematangan, dan tingkat kesegaran secara terpadu. penelitian dilakukan terhadap sepuluh jenis buah lokal indonesia dengan produksi tertinggi tahun 2024, yaitu pisang, mangga, salak, nanas, jeruk siam, durian, pepaya, alpukat, rambutan, dan nangka. model dirancang ringan dengan memanfaatkan teknik dual convolution, dua puluh lapisan backbone, serta tiga output head, menghasilkan total 41.296 parameter dengan ukuran memori 161,31 kb. hasil uji menunjukkan model yolo-cnn-mtl mencapai f1-score sebesar 86,49% untuk klasifikasi jenis buah, 96,11% untuk kesegaran, dan 87,55% untuk kematangan, lebih unggul dibandingkan pendekatan yolo tunggal. pendekatan mtl juga menunjukkan kemampuan generalisasi yang lebih baik dan dapat mempercepat proses inferensi hingga 2,35 kali dibandingkan single-task learning. hasil penelitian membuktikan pendekatan multi-task learning efektif dan efisien untuk mendukung sistem panen dan pascapanen buah berbasis deep learning, khususnya pada lingkungan dengan keterbatasan sumber daya komputasi.

Tropical fruits are major horticultural commodities in Indonesia that require proper harvesting and post-harvest handling to maintain their quality, maturity, and freshness. Fruit quality control processes are generally carried out manually due to limited computational resources, making them prone to errors, especially under diverse field conditions. This study was conducted to develop a Multi-Task Learning (MTL) model based on a hybrid approach of YOLOv12 and Convolutional Neural Networks (CNN) to detect and classify fruit types, maturity levels, and freshness in an integrated manner. The study was conducted on ten types of local Indonesian fruits with the highest production in 2024, namely bananas, mangoes, salak, pineapples, Siam oranges, durians, papayas, avocados, rambutans, and jackfruits. The model was designed to be lightweight by utilizing dual convolution techniques, twenty layers of backbone, and three output heads, resulting in a total of 41,296 parameters with a memory size of 161.31 KB. The test results show the YOLO-CNN-MTL model achieved F1-Scores of 86.49% for fruit type classification, 96.11% for freshness, and 87.55% for maturity, outperforming the single YOLO approach. The MTL approach also demonstrated better generalization capability and was able to accelerate the inference process by up to 2.35 times compared to Single-Task Learning. The results of this study prove that the Multi-Task Learning approach is effective and efficient in supporting deep learning-based fruit harvesting and post-harvest systems, particularly in environments with limited computational resources.