A detection algorithm based on improved YOLOv5 for coarse-fine variety fruits

• Published in: Journal of food measurement & characterization Vol. 18; no. 2; pp. 1338 - 1354
• Main Authors: Fan, Yanru, Cai, Yuliang, Yang, Huijun
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.02.2024; Springer Nature B.V
• Subjects: Accuracy; Agriculture; Algorithms; Chemistry; Chemistry and Materials Science; Chemistry/Food Science; Complexity; Convolution; Engineering; Food Science; Fruits; Mathematical models; Object recognition; orchards; Original Paper; Parameters; Precision agriculture; Real time; Sorting algorithms; species; Deep learning; Fruit detection; Coarse-fine variety fruits; Improved YOLOv5
• ISSN: 2193-4126; 2193-4134
• DOI: 10.1007/s11694-023-02274-z

Fruit detection and recognition is a key technology in precision agriculture such as automated picking, orchard yield measurement, and fruit sorting. Although current algorithms have good detection performance for single-class objects in living scenes, they have lower detection and recognition accuracy for different varieties of fruits with high similarity in complex environments and consume high computing resources and costs, which cannot be applied to edge devices for real-time detection and sorting of varieties fruits species. This paper proposed a lightweight model based on YOLOv5 for the detection and recognition of coarse-fine variety fruits in clean or complex scenes. First, the networks with different widths and depths of YOLOv5 were trained to find the best baseline detection net; second, GhostConv, a lightweight convolution layer, was introduced to reduce parameters and computations in the baseline network; finally, the input channels of the depth convolution in the backbone was expanded and the Wise-IoU bounding box loss function was introduced to improve the detection accuracy of the baseline network. The experimental results showed that, considering both detection performance and model complexity, YOLOv5s performs better as the baseline network. After optimization, the model parameters and weight volume were reduced by approximately 33%, the mean average precision at 0.5 IoU(mAP@.5) increased by 0.6%, and increased by 1.2% at IoU = 0.5:0.95(mAP@.5:.95). The improved model achieved the reasonable balance between detection accuracy and complexity. It is suitable for real-time detection, online grading, and rapid sorting of fruits in precision agriculture.