摘要
针对农场环境下使用手工标记与肉眼识别方法识别奶牛时计数效率低、错误率高,而使用无线射频检测技术比较复杂且成本高的问题,使用检测速度较快且性能较好的YOLOv3算法对农场环境下的奶牛进行目标识别。该方法采用了3个不同尺度的特征图来进行对象检测,能够检测到更加细粒度的特征;使用Darknet-53网络加入残差模块,有利于解决深层次网络的梯度问题,从而增加奶牛目标识别模型的识别效果;采用K-means聚类得到先验框的尺寸,预测对象类别时使用logistic的输出进行预测,可以支持多标签对象。从检测结果来看,该方法在农场环境背景下的奶牛目标识别效果较好,检测准确率较高。
The method of manual marking and visual recognition for the detection of cows in stockbreeding is low in recognition efficiency and high in error rate, and the use of radio frequency detection technology is complicated and costly. The object detection of cows based on the YOLOv3 algorithm is fast in detection speed and good in performance. This method uses three different scale feature maps for object detection, which can detect more fine-grained features of cow images. Using the darknet-53 network to join the residual module is beneficial to solving the gradient problem of deep network, thus increasing the effect of the cow target detection model;the size of the priori box is obtained by K- means clustering, and the logistic output is used for prediction when predicting the object class, which can support multi-label objects. Test results show that the method has better detection effect on dairy cows in farm environment, and the detection accuracy is higher.
作者
王毅恒
许德章
WANG Yiheng;XU Dezhang(College of Mechanical and Automotive Engineering, Anhui Engineering University, Wuhu 241000, China;Wuhu Anpu Robot Technology Institute Co. Ltd., Wuhu 241007, China)
出处
《广东石油化工学院学报》
2019年第4期31-35,共5页
Journal of Guangdong University of Petrochemical Technology
基金
国家自然科学基金项目(61741101)
安徽省高校自然科学研究重大项目(KJ2018ZD014)
安徽省重点研发计划(1804a09020036)
安徽省科技攻关项目(1604a0902125)
安徽省自然科学基金项目(1608085QF154)
