摘要
为了实现对复杂养殖场环境下奶牛个体的精准识别,针对SSD(single shot multibox detector)算法对重叠对象检测效果不好的问题,对SSD算法进行改进。通过将不同特征图的特征进行融合,使得不同特征图的信息可以互补,从而改善算法对重叠对象的检测效果;去掉网络中的Conv4_3层特征图,同时增加其他特征图候选框的数量,这样不仅可以保证算法的实时性,而且提高了检测精度;引入迁移学习方法,以提高算法的平均准确率。实验结果表明:改进的SSD算法与传统SSD算法相比,在满足实时检测的情况下,平均准确率(AP)提高4.32%;经过迁移后,改进SSD算法的AP提高3.85%。
To realize accurate identification of individual cows in a complex farm environment,the SSD(single shot multibox detector)algorithm was improved to solve the problem of poor detection effect for overlapping objects.First,different feature maps were fused to ensure that different feature maps complemented each other and improved the detection effect of overlapping objects.Then,Conv4_3 was removed from the network.The number of candidate frames in other feature maps increased,ensuring the real-time performance of the algorithm and also improving the detection accuracy.Finally,the transfer learning method was used to improve the average accuracy of the algorithm.The experimental results show that compared with the traditional SSD algorithm,the average accuracy(AP)of the improved SSD algorithm is improved by 4.32%in real-time detection,and the AP of the improved SSD algorithm is increased by 3.85%after migration.
作者
邢永鑫
吴碧巧
吴松平
王天一
Xing Yongxin;Wu Biqiao;Wu Songping;Wang Tianyi(College of Big Data and Information Engineering,Guizhou University,Guiyang 550025,Gruizhou,China)
出处
《激光与光电子学进展》
CSCD
北大核心
2021年第16期495-503,共9页
Laser & Optoelectronics Progress
基金
贵州省科技支撑计划(SY[2017]2881)。
关键词
图像处理
人工智能
目标识别
卷积神经网络
特征融合
迁移学习
image processing
artificial intelligence
target recognition
convolutional neural network
feature fusion
transfer learning
