基于多通道注意力机制的密集草莓成熟度识别神经网络

曹昭睿¹，王楚文¹，辛晓杰¹，陈心垚¹，才佳卉^2*

(1. 沈阳理工大学装备工程学院，沈阳110159，中国；

2. 渤海大学食品科学与工程学院，锦州121013，辽宁，中国)

摘要：颜色是草莓成熟度的重要指标，因此识别采摘和运输过程中的颜色变化至关重要。草莓柔软的外皮在采摘和人工挑选过程中容易造成机械损伤。尽管学者们在利用机器视觉技术进行作物检测方面取得了显著进展，但在检测密集放置、大小不一且处于不同生长状态的草莓时，仍然面临挑战。该研究提出了一种基于YOLOv8s的密集草莓成熟度识别卷积神经网络，该网络集成了多尺度特征注意力机制。

所提出的模型解决了多个问题，包括相邻水果特征的重叠、模糊的成熟度标准，以及在密集放置的草莓中难以区分单个水果的问题。该模型利用多通道注意力机制融合不同尺度草莓的语义特征，增强了在密集且随机放置的环境中单个草莓图像信息的独立概括能力。此外，该模型还引入了小波下采样卷积作为网络层的骨干，以提高捕捉小草莓详细特征的能力。此外，通过集成加权交并损失函数，优化了网络训练的收敛效果和推理准确性。

在自定义的草莓数据集上，与原始YOLOv8s模型相比，模型的准确率、召回率、mAP@0.5和mAP@0.5:0.95分别提高了1.1%、1.8%、1.2%和0.8%，表明该模型在面对多种大小和任意放置的密集草莓阵列时，具有良好的识别准确性和稳定性。所提出的模型能够在随机环境中快速检测单个草莓果实的成熟度，提高分选效率，并减少采后损失。该研究为机器视觉技术在密集作物成熟度识别领域的应用提供了新的思路和技术参考。

关键词：目标识别；特征融合；成熟度检测；深度学习

DOI: 10.25165/j.ijabe.20261901.9931

引用信息: Cao Z R, Wang C W, Xin X J, Chen X Y, Cai J H. Dense strawberry maturity recognition neural network based on multichannel attention mechanism. Int J Agric & Biol Eng, 2026; 19(1): 295–301.

Dense strawberry maturity recognition neural network based on multichannel attention mechanism

Zhaorui Cao¹, Chunwen Wang¹, Xiaojie Xin¹, Xinyao Chen¹, Jiahui Cai^2*

(1. College of Equipment Engineering, Shenyang Ligong University, Shenyang 110159, China; 

2. College of Food Science and Engineering, Bohai University, Jinzhou 121013, Liaoning, China)

Abstract: Color is an important indicator of strawberry maturity; therefore, identifying color changes during harvesting and delivery is important. The soft skin of strawberries can easily cause mechanical damage during harvesting and manual selection. While scholars have made significant progress in using machine vision technology for crop detection, detecting strawberries that are densely placed, differently sized, and in different growth states remains challenging. This study proposes a YOLOv8s-based dense strawberry maturity recognition convolutional neural network that integrates multiscale feature attention. The proposed model addresses several issues, including the overlapping of adjacent fruit features, fuzzy maturity criteria, and difficulty distinguishing individual fruits when they are densely placed. It utilizes a multichannel attention mechanism to fuse semantic features of strawberries at different scales, enhancing the independent summarization ability of image information of individual strawberries in dense and randomly placed environments. It also introduces wavelet down sampling convolution as the backbone of network layers to enhance the ability to capture detailed features of small strawberries. Furthermore, with the integration of the weighted intersection over union loss function, it optimizes the convergence effect and inference accuracy of network training. On a custom strawberry dataset, model accuracy, recall rate, mAP@0.5, and mAP@0.5:0.95 increased by 1.1%, 1.8%, 1.2%, and 0.8%, respectively, compared to the original YOLOv8s model, showing good recognition accuracy and stability when facing dense strawberry arrays with multiple sizes and arbitrary placement. The proposed model can quickly detect the maturity of individual strawberry fruits in random environments, improve sorting efficiency, and reduce post-harvest losses. This study provides new ideas and technical references for the application of machine vision technology in the field of dense crop maturity recognition.

Keywords: object detection, feature fusion, maturity testing, deep learning

DOI: 10.25165/j.ijabe.20261901.9931

Citation: Cao Z R, Wang C W, Xin X J, Chen X Y, Cai J H. Dense strawberry maturity recognition neural network based on multichannel attention mechanism. Int J Agric & Biol Eng, 2026; 19(1): 295–301.

期刊官网：https://www.ijabe.org      http://www.ijabe.net

联系电话：010-59197091

联系邮箱：ijabe@ijabe.cn

支持单位：