基于改进全卷积网络的临产母猪体态识别  

Posture Recognition of Parturient Sow Based on Improved Fully Convolutional Networks

作　　者：王萍[1] 侯岩松 许海洋[1] 张伟[2] WANG Ping;HOU Yansong;XU Haiyang;ZHANG Wei(School of Science and Information,Qingdao Agricultural University,Qingdao,Shandong 264300;School of Applied Technology,China University of Labor Relations,Beijing 100048,China)

机构地区：[1]青岛农业大学理学与信息科学学院,山东青岛264300 [2]中国劳动关系学院应用技术学院,北京100048

出　　处：《贵州农业科学》2023年第6期120-127,共8页Guizhou Agricultural Sciences

基　　金：科技创新2030新一代人工智能重大项目“典型畜禽疫病知识图谱构建与演化”(2021ZD0113802);教育部第二批新工科研究与实践项目“工会特色高校多主体协同新工科人才培养实践创新平台建设探索与实践”(E-XTYR20200608);教育部产学合作协同育人项目“基于Android的茶叶害虫识别系统”(201801006084、202102381007);青岛农业大学高层次人才科研基金资助项目“基于群体智能的变异测试用例生成方法”(663/1117028)。

摘　　要：【目的】优化临产母猪体态识别的方法,为客观反映母猪生理状态和有效预防母猪疾病提供参考。【方法】基于改进全卷积网络(FCN)的母猪体态识别算法,构建FCN训练模型,采用自适应学习率进行训练,识别临产母猪体态。【结果】模型训练前15轮模型loss和评价指标mae迅速下降,30轮时模型基本完全收敛,且模型在训练集和测试集上的准确率可达100%;模型微调后,loss和mae仍有小范围下降,但对模型最终预测结果的提升不明显。不同光照条件下该模型可有效识别定位母猪的不同体态,具有较强鲁棒性,模型平均识别准确率为97%,且能较好地捕捉样本特征,预测框IOU为92.56%,运行速度为0.2 s,基本可达实时检测要求。【结论】基于改进全卷积网络的临产母猪体态识别方法可在光照条件欠佳的猪舍环境中有效识别母猪体态,实现对临产母猪的实时监测。【Objective】The method of posture recognition of parturient sow is optimized,which provides reference for reflecting physiological state of sows objectively and preventing disease of sows effectively.【Method】Based on the improved full convolutional network(FCN)algorithm,the FCN training model is constructed,and the self-adaptive learning rate is used for training to identify the posture of parturient sow.【Result】In the first 15 rounds of model training,model loss and evaluation metrics“mae”decreased rapidly,and the model was almost completely convergent at 30 rounds,and the accuracy of the model on the training set and the test set could reach 100%.After the fine-tuning of the model,both loss and mae decreased to a certain extent,but the improvement of the final prediction result of the model was not obvious.The model can effectively identify and locate different posture of sows under different lighting conditions,and has strong robustness.The average recognition accuracy of the model is 97%,and the sample characteristics can be captured well.The prediction frame IOU is 92.56%,and the running speed is 0.2 seconds,which can basically meet the real-time detection requirements.【Conclusion】This method can effectively identify the posture of sows in the poor lighting environment and realize real-time monitoring of parturient sow.

关 键 词：临产母猪 体态识别 注意力机制 全卷积网络 迁移学习 

分 类 号：S828[农业科学—畜牧学] TP391[农业科学—畜牧兽医]