基于深度学习的牛脸目标检测研究  

Research on Cow face Object Detection Based on Deep Learning

作　　者：郭祯霖 石建飞[1] 佟柏宏 吴继祯 Guo Zhenlin;Shi Jianfei;Tong Baihong;Wu Jizhen(Heilongjiang Bayi Agricultural University,Daqing 163319)

机构地区：[1]黑龙江八一农垦大学,大庆163319

出　　处：《黑龙江八一农垦大学学报》2024年第4期86-97,共12页journal of heilongjiang bayi agricultural university

基　　金：黑龙江八一农垦大学自然科学人才支持计划(ZRCPY201814)。

摘　　要：为解决牛脸目标识别精度问题,研究以深度学习图像处理技术为支撑,采用一种基于轻量型牛脸数据集训练下的Yolov5目标检测算法模型,对采集到的涵盖复杂背景的牛只图像中的牛脸目标进行识别。在Yolov5模型基础上针对牛脸部分细小目标对象检测能力做出改进。引入CBAM即插即用的注意力机制,增强网络对有意义区域的感知能力,并减少对牛舍背景复杂环境噪声等干扰信息的影响。融合BiFPN加权双向特征金字塔网络结构,可有效地融合牛只个体面部深层和浅层特征,提高网络对图像中包含牛只个体中大、小牛脸面部目标的检测能力。研究以小样本牛脸数据集支持,牛脸目标检测的平均准确率为0.934。结果表明,该研究可在实际生产中对牛脸目标进行有效检测。In order to solve the problem of cow face target recognition accuracy,a Yolov5 target detection algorithm model based on the training of lightweight bovine face data set was adopted to recognize bovine face targets in the acquired cow images covering complex backgrounds,supported by deep learning image processing technology.Based on Yolov5 model,the detection ability of small bovine face objects was improved.CBAM plug and play attention mechanism was introduced to enhance the network's ability to perceive meaningful areas and reduce the influence of interference information such as complex environmental noise in the background of cow barn.The fusion of BiFPN weighted bidirectional feature pyramid network structure could effectively merge the deep and shallow features of individual bovine faces,and improve the detection ability of the network for objects containing large and small cow faces in images.The average accuracy of bovine face detection was O.934,supported by small sample cow face data set.The results showed that this research could effectively detect the cow face target in actual production.

关 键 词：深度学习 目标检测 牛脸检测 Yolov5s 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]