用于体质量估测的黄羽鸡姿态关键帧识别与分析

Posture Key Frame Recognition and Analysis for Weight Estimation of Yellow-feathered Chickens

体质量是评价家禽生长状况的关键指标,但家禽姿态的变化会影响体质量估测精度。本研究提出了一种SE-ResNet18+fLoss网络对平养模式下黄羽鸡姿态关键帧进行识别,融合了注意力机制SE模块和残差结构,并改进了损失函数,通过Focal Loss监督信号来解决样本不平衡问题,同时引入梯度加权类激活图对末端分类规则的合理性进行解释。利用4295幅鸡只图像构建数据集,测试集中鸡只的站立、低头、展翅、梳理羽毛、坐姿和遮挡6类姿态情况识别的F1值分别为94.34%、91.98%、76.92%、93.75%、100%和93.68%;黄羽鸡姿态关键帧的识别精确率为97.38%、召回率为97.22%、F1值为97.26%、识别速度为19.84 f/s,识别精度、召回率和F1值均优于ResNet18、MobileNet18 V2和SE-ResNet18网络,在提高黄羽鸡姿态关键帧识别精度的同时保证了实时性,为准确估测家禽体质量提供了技术支持。

Body weight is a key indicator to evaluate the growth condition of poultry.However,the variation of poultry posture will affect the accuracy of weight estimation.SE-ResNet18+fLoss network was proposed to detect the posture key frames of floor-reared yellow-feathered chickens.The attention mechanism SE module and residual structure were integrated.And the Focal Loss was added to solve the problem of sample imbalance.In addition,the Gradient-weighted Class Activation Mapping was introduced to explain the rationality of the end classification rule.The dataset was constructed by 4295 images of yellow-feathered chickens.The F1-score of the SE-ResNet18+fLoss model on the test set for the chicken situations recognition of six classes:standing,bowing head,spreading wing,grooming feather,sitting and occlusion were 94.34%,91.98%,76.92%,93.75%,100%and 93.68%,respectively.Towards the detection of key posture frames on chickens,the accuracy,recall,F1-score and detection speed were 97.38%,97.22%97.26%and 19.84 f/s,respectively.And the detection accuracy,recall and F1-score were better than those of ResNet18,MobileNet V2 and SE-ResNet18 networks.The study ensured real-time performance while improving the accuracy of key posture frame recognition,which provided technical support for accurate estimation of poultry weight.