首页|PyTorch框架下的复杂场景目标识别方法研究

PyTorch框架下的复杂场景目标识别方法研究

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
人工智能框架(PyTorch+OneDNN)在复杂场景目标识别上,易陷入网络参数梯度超规模的问题.为此,提出一种基于PyTorch框架的复杂场景目标识别方法.引入MPI中的Ring Allreduce算法,优化PyTorch框架,以此实现复杂场景特征迭代提取过程中的超规模数据同步和规约处理.以优化后的PyTorch框架为基础,考虑到复杂场景目标特征与背景特征之间的交叉性,构建增强多尺度特征层输出的目标特征之间的关联.借助反卷积特征融合操作和残差融合操作的优势,依据上述关联性,实现目标自动识别.测试结果表明:所提方法的整体错误识别数量为113个,整体未识别数量为107个,证明了所提方法具有较优的自动化识别效果.
Research on object recognition methods in complex scenes under the PyTorch framework
The artificial intelligence framework(PyTorch+OneDNN)is prone to the problem of network parameter gradient su-perscale in complex scene target recognition.To this end,a complex scene object recognition method based on the PyTorch frame-work is proposed.Introducing the Ring Allreduce algorithm in MPI,optimizing the PyTorch framework to achieve super scale data synchronization and reduction processing in the iterative feature extraction process of complex scenes.Based on the optimized Py-Torch framework,considering the intersection between complex scene target features and background features,we construct an en-hanced correlation between target features output by multi-scale feature layers.By leveraging the advantages of deconvolution fea-ture fusion and residual fusion operations,and based on the above correlation,automatic target recognition can be achieved.The test results show that the overall number of errors identified by the proposed method is 113,and the overall number of unrecognized errors is 107,proving that the proposed method has better automation recognition performance.

PyTorch frameworkcomplex scenetarget automation identificationRing Allreduce algorithmScatter Reduce operationAllGather operation

张进军

展开 >

安徽警官职业学院信息管理系,合肥 230031

PyTorch框架 复杂场景 目标自动化识别 Ring Allreduce算法 Scatter Reduce操作 AllGather操作

安徽省高等学校自然科学研究重点项目(2023)

2023AH052757

2024

10.3969/j.issn.1007-1423.2024.08.011

现代计算机
中大控股

现代计算机

影响因子：0.292
ISSN：1007-1423
年,卷(期)：2024.30(8)
  • 12