Fast and accurate detection of kiwifruit in orchard using improved YOLOv3-tiny model

• Published in: Precision agriculture Vol. 22; no. 3; pp. 754 - 776
• Main Authors: Fu, Longsheng, Feng, Yali, Wu, Jingzhu, Liu, Zhihao, Gao, Fangfang, Majeed, Yaqoob, Al-Mallahi, Ahmad, Zhang, Qin, Li, Rui, Cui, Yongjie
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.06.2021; Springer Nature B.V
• Subjects: Actinidia; Agriculture; Algorithms; Atmospheric Sciences; automatic detection; Biomedical and Life Sciences; Chemistry and Earth Sciences; color; Complexity; Computer applications; Computer Science; Convolution; Deep learning; Fruits; Illumination; Kiwifruit; Life Sciences; lighting; Machine learning; Night; Object recognition; Orchards; Physics; precision agriculture; Remote Sensing/Photogrammetry; Semantics; Soil Science & Conservation; Statistics for Engineering; Deep learning; Data augmentation; YOLOv3-tiny model; Convolutional kernel; Image detection
• ISSN: 1385-2256; 1573-1618
• DOI: 10.1007/s11119-020-09754-y

Automatic detection of kiwifruit in the orchard is challenging because illumination varies through the day and night and because of color similarity between kiwifruit and the complex background of leaves, branches and stems. Also, kiwifruits grow in clusters, which may result in having occluded and touching fruits. A fast and accurate object detection algorithm was developed to automatically detect kiwifruits in the orchard by improving the YOLOv3-tiny model. Based on the characteristics of kiwifruit images, two convolutional kernels of 3 × 3 and 1 × 1 were added to the fifth and sixth convolution layers of the YOLOv3-tiny model, respectively, to develop a deep YOLOv3-tiny (DY3TNet) model. It takes multiple 1 × 1 convolutional layers in intermediate layers of the network to reduce the computational complexity. Testing images captured from day and night and comparing with other deep learning models, namely, Faster R-CNN with ZFNet, Faster R-CNN with VGG16, YOLOv2 and YOLOv3-tiny, the DY3TNet model achieved the highest average precision of 0.9005 with the smallest data weight of 27 MB. Furthermore, it took only 34 ms on average to process an image of a resolution of 2352 × 1568 pixels. The DY3TNet model, along with the YOLOv3-tiny model, showed better performance on images captured with flash than those without. Moreover, the experiments indicated that the image augmentation process could improve the detection performance, and a simple lighting arrangement could improve the success rate of detection in the orchard. The experimental results demonstrated that the improved DY3TNet model is small and efficient and that it would increase the applicability of real-time kiwifruit detection in the orchard even when small hardware devices are used.