摘要
鉴于目标检测中的物体外观会根据其基本形状及不同的姿势和视角而有很大的差异,对Faster R-CNN算法进行研究并提出一种多通道检测算法。根据图像宽高比给生成的RoI分配由3个通道组成的网络进行训练和测试,通过最小化正则函数R(W)和3对损失函数之和L(W)来优化网络,3个通道共享fc6层来提高检测性能并节省内存空间。为验证算法的有效性,在多个数据集和自己拍摄的图像上进行实验验证,实验结果表明,在PASCALVOC2012数据集中改进算法平均精度为78.8%,相比其它相关算法在不同程度上有所提高。
In view of the fact that the appearance of objects in object detection will vary greatly according to their basic shapes and different poses and perspectives,the Faster R-CNN algorithm was studied and a multi-channel detection algorithm was proposed.According to the aspect ratio of the image,the generated RoI was assigned a network consisting of three channels for training and testing.The network was optimized by minimizing the regular function R(W)and the sum of three pairs of loss functions L(W).The three channels shared the fc6 layer to improve detection performance and save memory space.To verify the effectiveness of the algorithm,experiments were carried out on multiple data sets and self-photographed images.Experimental results show that the ave-rage accuracy of the improved algorithm in the PASCALVOC2012 data set is 78.8%,which is improved to varying degrees compared with other related algorithms.
作者
殷小芳
辛月兰
兰天
何晓明
YIN Xiao-fang;XIN Yue-lan;LAN Tian;HE Xiao-ming(College of Physics and Electronic Information Engineering,Qinghai Normal University,Xining 810000,China)
出处
《计算机工程与设计》
北大核心
2021年第12期3453-3460,共8页
Computer Engineering and Design
基金
国家自然科学基金项目(61662062)
青海省重大科技专项子课题基金项目(2019-ZJ-A10)
青海省科技厅基础研究计划基金项目(2018-ZJ-719)。
