Sui Jinxue, Liu Li, Wang Zuoxun, Yang Li
School of Information and Electronic Engineering, Shandong Technology and Business University, Yantai, China.
PLoS One. 2025 Mar 3;20(3):e0319041. doi: 10.1371/journal.pone.0319041. eCollection 2025.
The widespread cultivation of apples highlights the importance of efficient and accurate apple detection algorithms in robotic picking technology. The accuracy of current apple picking detection algorithms is still limited when the distribution is dense and occlusion exists, and there is a significant challenge in deploying current high accuracy detection models on edge devices with limited computational resources. To solve the above problems, this paper proposes an improved detection algorithm (RE-YOLO) based on YOLOv8n. First, this paper innovatively introduces Receptive-Field Attention Convolution (RFAConv) to improve the backbone and neck network of YOLOv8. It essentially solves the problem of convolution kernel parameter sharing and improves the consideration of the differential information from different locations, which significantly improves the accuracy of model recognition. Second, this paper innovatively proposes an EMA_C2f module. This module makes the spatial semantic features uniformly distributed to each feature group through partial channel reconstruction and feature grouping, which emphasizes the interaction of spatial channels, improves the ability to detect subtle differences, can effectively discriminate the apple occlusion, and reduces the computational cost. Finally, the loss function of YOLOv8 is improved using the Wise Intersection over Union (WIOU) function, which not only simplifies the gradient gain assignment mechanism and improves the ability to detect targets of different sizes, but also accelerates the model optimization. The experimental results show that RE-YOLO improves the precision, recall, mAP@0.5, and mAP@0.5-0.95 by 2%, 2.1%, 2.7%, and 3.9%, respectively, compared with the original YOLOv8. Compared with YOLOv5, it improves 4%, 1.9%, 1.7% and 3%, respectively, which fully proves the advanced and practical nature of the proposed algorithm.
苹果的广泛种植凸显了高效准确的苹果检测算法在机器人采摘技术中的重要性。当苹果分布密集且存在遮挡时，当前苹果采摘检测算法的准确性仍然有限，并且在计算资源有限的边缘设备上部署当前的高精度检测模型存在重大挑战。为了解决上述问题，本文提出了一种基于YOLOv8n的改进检测算法（RE-YOLO）。首先，本文创新性地引入了感受野注意力卷积（RFAConv）来改进YOLOv8的主干和颈部网络。它本质上解决了卷积核参数共享的问题，并改进了对来自不同位置的差异信息的考量，显著提高了模型识别的准确性。其次，本文创新性地提出了一个EMA_C2f模块。该模块通过部分通道重建和特征分组使空间语义特征均匀分布到每个特征组，强调了空间通道的交互作用，提高了检测细微差异的能力，能够有效区分苹果的遮挡情况，并降低了计算成本。最后，使用明智交并比（WIOU）函数改进了YOLOv8的损失函数，这不仅简化了梯度增益分配机制，提高了检测不同大小目标的能力，还加速了模型优化。实验结果表明，与原始的YOLOv8相比，RE-YOLO的精度、召回率、mAP@0.5和mAP@0.5 - 0.95分别提高了2%、2.1%、2.7%和3.9%。与YOLOv5相比，它分别提高了4%、1.9%、1.7%和3%，充分证明了所提算法的先进性和实用性。
