Fruit automatic recognition system base on YOLO algorithm
Guo, Zitao (2025)
Kandidaatintyö
2025
School of Engineering Science, Tietotekniikka
Kaikki oikeudet pidätetään.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2025050838385
https://urn.fi/URN:NBN:fi-fe2025050838385
Tiivistelmä
With the advancement of deep learning computer vision, real-time fruit detection has become increasingly essential in smart agriculture and supply chain management. High-precision fruit recognition not only improves productivity but also reduces errors associated with traditional manual identification, thereby helping enterprises reduce waste and improve operational efficiency.
This study proposes an automatic fruit recognition system based on the YOLOv5 (You Only Look Once version 5), which performs real-time detection and identification of apples, bananas, and oranges in natural images. The system is trained on a custom-labeled dataset from Kaggle, Encompassing 240 training images and 60 test images. Data augmentation techniques and pre-trained COCO weights are employed to enhance the model’s generalization capability. The system achieves excellent performance in terms of precision, recall, F1 score, and mean average precision (mAP), while supporting real-time detection at 40 frames per second (FPS).
A convenient web interface is developed to ensure effective human-computer interaction. Experimental results demonstrate that the proposed system has strong potential for deployment in practical scenarios such as smart orchards, food packaging lines, and retail surveillance environments.
This study proposes an automatic fruit recognition system based on the YOLOv5 (You Only Look Once version 5), which performs real-time detection and identification of apples, bananas, and oranges in natural images. The system is trained on a custom-labeled dataset from Kaggle, Encompassing 240 training images and 60 test images. Data augmentation techniques and pre-trained COCO weights are employed to enhance the model’s generalization capability. The system achieves excellent performance in terms of precision, recall, F1 score, and mean average precision (mAP), while supporting real-time detection at 40 frames per second (FPS).
A convenient web interface is developed to ensure effective human-computer interaction. Experimental results demonstrate that the proposed system has strong potential for deployment in practical scenarios such as smart orchards, food packaging lines, and retail surveillance environments.