作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2018, Vol. 44 ›› Issue (12): 222-227. doi: 10.19678/j.issn.1000-3428.0049051

• 人工智能及识别技术 • 上一篇    下一篇

基于混合结构卷积神经网络的目标快速检测算法

林封笑,陈华杰,姚勤炜,张杰豪   

  1. 杭州电子科技大学 自动化学院,杭州 310018
  • 收稿日期:2017-10-23 出版日期:2018-12-15 发布日期:2018-12-15
  • 作者简介:林封笑(1993—),男,硕士研究生,主研方向为模式识别、机器学习;陈华杰,教授、博士;姚勤炜、张杰豪,硕士研究生

Target Fast Detection Algorithm Based on Hybrid Structure Convolutional Neural Network

LIN Fengxiao,CHEN Huajie,YAO Qinwei,ZHANG Jiehao   

  1. Automated Institute,Hangzhou Dianzi University,Hangzhou 310018,China
  • Received:2017-10-23 Online:2018-12-15 Published:2018-12-15

摘要:

为提高基于卷积神经网络(CNN)目标检测算法的检测速度,提出一种基于混合结构CNN的目标快速检测算法。采用基于CNN的Faster R-CNN目标检测框架,对其CNN进行优化。基于多层感知器结构,提出CR-mlpconv卷积层结构。在网络浅层采用C.ReLU策略,同时结合CR-mlpconv层结构和C.ReLU策略,合理设计层参数,构成卷积神经网络。将该卷积神经网络融合到Faster R-CNN检测框架中,实现目标快速检测。实验结果表明,在检测精度的适当影响范围内,该算法能够减少网络模型参数并降低网络模型的内存消耗,提高网络的实时性。

关键词: 目标快速检测, Faster R-CNN框架, 卷积神经网络, 特征提取, 混合结构, 低通道

Abstract:

In order to improve the detection speed of the Convolutional Neural Network (CNN) target detection algorithm,a target fast detection algorithm based on hybrid structure CNN is proposed.CNN is optimized by CNN-based Faster R-CNN target detection framework.Based on the multilayer perceptron structure,a CR-mlpconv convolutional layer structure is proposed.The C.ReLU strategy is adopted in the shallow layer of the network,and the CR-mlpconv layer structure and the C.ReLU strategy are combined to design the layer parameters reasonably to form CNN.CNN is merged into the Faster R-CNN detection framework to achieve rapid target detection.Experimental results show that compared with the Faster R-CNN+ZFnet algorithm,the algorithm can reduce the network model parameters,reduce the memory consumption of the network model,and improve the real-time performance of the network.

Key words: target fast detection, Faster R-CNN framework, Convolutional Neural Network(CNN), feature extraction, hybrid structure, low channel

中图分类号: