基于红外热成像技术的笼内死鸡自动识别方法  

Automatic identification method for dead chicken in cage based on infrared thermal imaging technology

作　　者：贾雁琳 薛皓 周子轩 赵学谦 霍晓静[1] 李丽华[1] JIA Yanlin;XUE Hao;ZHOU Zixuan;ZHAO Xueqian;HUO Xiaojing;LI Lihua(College of Mechanical and Electrical Engineering,Hebei Agricultural University,Baoding 071001,China)

机构地区：[1]河北农业大学机电工程学院,河北保定071001

出　　处：《河北农业大学学报》2023年第3期105-112,共8页Journal of Hebei Agricultural University

基　　金：国家自然科学基金项目(31902209);河北省重点研发计划项目(20327220D,20326630D);河北省现代农业产业创新团队岗位科学家(HBCT2018060204);河北省省级科技计划资助项目(22326607D)。

摘　　要：目前集约化养鸡场主要采用层叠式立体笼养模式,进行死鸡巡检过程,工人需多次攀爬扶梯,劳动强度大且工作简单机械重复。为了提高劳动效率,增加人工智能对劳动型人才进行补充,本文将图像识别分析与红外热成像技术相结合,采用了利用温度阈值剥离出鸡头特征再提取形态学特征,结合支持向量机的死鸡识别方法。首先对图像预处理,对应红外温度(T)—灰度值(G)线性函数,找到鸡头与背景的剥离温度阈值;计算出红外热像中鸡头平均灰度值,与设定的剥离阈值比较,保存图片中标记的样本目标。提取鸡头样本的形态特征向量,并基于XGBoost进行特征筛选,选择得分排序前五的圆形度R、紧凑度J、离心率E、长轴长L、短轴长S作为分类特征向量,最后利用支持向量机分类器实现活鸡与死鸡的区分。实验结果表明:在对比分类模型效果时,基于决策树算法的死鸡分类准确率为87.5%,基于BP神经网络和支持向量机算法的分类准确率为91.67%,其中支持向量机分类算法召回率、F1和AUC的值最高。可为实现多层立体笼养死禽自动识别提供1种新方法。At present,the intensive chicken farm mainly adopts the layered three-dimensional cage mode to inspect the dead chickens.Workers need to climb the escalator many times to screen out 10-15 dead chickens from nearly 30000 chickens in the cages,which takes at least one and a half hours.The labor intensity is high,but the work is simple and repetitive.In order to improve labor efficiency and increase artificial intelligence to supplement labor-oriented talents,this paper combined image recognition analysis with infrared thermal imaging technology,and proposed a method for dead chicken recognition.The chicken head feature was stripped by temperature threshold followed by extraction of the morphological feature combined with support vector machine.The image was preprocessed according to the linear function of infrared temperature(T)-gray value(G)to determine the stripping temperature threshold of chicken head from background.The average gray value of the chicken head in the infrared thermal image was calculated and compared with the set stripping threshold to save the sample target marked in the image.The morphological feature vectors of chicken head samples were then extracted and the features were screened based on XGBoost.The top five factors of score ranking,such as roundness R,compactness J,eccentricity E,long axis length L,short axis length S,were selected as the classification feature vectors.Finally,the support vector machine classifier was used to distinguish dead chicken from live chicken.The experimental results showed that the classification accuracy of dead chicken based on decision tree algorithm was 87.5%,while the classification accuracy based on BP neural network and support vector machine was 91.67%.The recall rate,F1 and AUC of support vector machine were the highest.This study provided a method for automatic recognition of dead birds in multi-layer three-dimensional cages.

关 键 词：红外热成像 图像识别 死鸡 多特征值 XGBoost SVM 

分 类 号：S24[农业科学—农业电气化与自动化]