轻量级多任务的苹果成熟度分类模型(特邀)

Lightweight Multi-Task Apple Ripeness Classification Model(Invited)

苹果成熟度等级和外观缺陷是判断其品质等级的两项重要依据。为实现采摘任务中未成熟、有缺陷苹果的自动剔除,提出一种轻量级多任务的成熟度分类模型(L-MTCNN)。该模型由D-Net和M-Net两个子网络构成,可实现苹果外观缺陷与成熟度等级分类的多任务判别。该模型利用一个主干网络提取特征信息,并将特征信息分别应用于D-Net和M-Net,提升模型特征的利用率,减少整体识别计算时间。引入Triplet损失作为M-Net的损失函数,增大不同等级间差距,减小相同等级内差距。此外,在行业标准的基础上,研究多种苹果成熟过程的外观变化,构建苹果成熟度数据集。针对图像采集过程中遇到的光线过强和光线过暗而造成所采集苹果图像与真实苹果外观颜色不一致的情况,提出一种基于亮度的颜色恢复算法,实现了采集图像的颜色恢复,并建立基于苹果成熟度的数据集。最终实验结果表明,D-Net、M-Net相较于AlexNet、ResNet18、ResNet34、VGG16在平均准确率方面有较大幅度的提升。此外,在召回率、精准率、F1分数方面,所提模型在成熟度等级、缺陷与否的分类任务上都具有更优的表现。所提模型可实现不同类型苹果的高准确率成熟度等级判断,为实现一体化作业机器人提供一定的借鉴和参考。

The maturity level and appearance defects of apples are crucial criteria for determining their quality.To automate the removal of immature and defective apples in picking tasks,a lightweight multi-task maturity classification model(L-MTCNN)is proposed.This model comprises two sub networks,D-Net and M-Net,for multi-task classification of apple appearance defects and maturity level.Furthermore,it uses a backbone network to extract feature information,which is then applied to D-Net and M-Net,thereby improving feature utilization and reducing overall recognition computation time.Introducing Triplet loss as the loss function for M-Net increases the separation between different maturity levels while reducing the variance within the same level.Additionally,based on industry standards,the study examines the appearance changes in various apple ripening processes and constructs an apple maturity dataset.A brightness-based color restoration algorithm is proposed to address the inconsistencies between collected apple images and their actual appearance,caused by varying lighting conditions during image acquisition.This algorithm restores the color restoration of the collected images and facilitates the creation of a reliable on apple maturity dataset.Experimental results indicate that D-Net and M-Net substantially improve average accuracy compared to AlexNet,ResNet18,ResNet34,and VGG16.Furthermore,in terms of recall rate,precision rate,and F1 score,the proposed model outperforms existing models in classifying maturity levels and defect statuses.This demonstrates that the model can achieve high-accuracy maturity level judgments for different types of apples,providing valuable insights for developing integrated operation robots.