增强层次CNN模型在目标识别应用中的研究  

Research on the Application of Intensive Hierarchical Convolution Neural Network Model in Target Recognition

作　　者：史天予 胡玉兰[1] 孙家民[1] 袁德鹏[2] 

机构地区：[1]沈阳理工大学信息科学与工程学院,沈阳110159 [2]东北大学计算机科学与工程学院,沈阳110819

出　　处：《光电技术应用》2016年第4期66-72,共7页Electro-Optic Technology Application

基　　金：国家自然科学基金项目(61373089)

摘　　要：受生物视觉信息处理机制启发的目标识别是当前计算机视觉领域研究的热点之一,其主要思想是对大脑视觉皮层中视觉信息的层次性处理过程进行模拟,构建数学模型来实现目标识别。然而传统的层次化计算模型通常以前馈信息传递为基础,层与层之间采用被动的硬连接方式,强调对视觉信息的多层分解,却较少涉及视觉神经系统的主动感知和学习过程。因此选择以同时具备稀疏连接思想和自我学习机制、并且具备良好网络拓扑结构的卷积神经网络为框架,基于经典卷积神经网络模型,融入分层和仿生的思想,提出新的基于视觉神经增强层次CNN模型——IH-CNN。实验结果表明,IH-CNN模型可以较好的解决大规模图像中的目标识别问题,目标识别准确率高达84%。Object recognition inspired by biological visual information processing mechanism is one of the re-search subjects in current computer vision research field, the main idea is to simulate the hierarchical process of vi-sual information in brain visual cortex and build mathematic model to achieve target recognition. However, tradition-al hierarchical calculation model is usually built based on front feed information transfer, and the passive hardwired way is used between layer and layer. The multi level decomposition of visual information is emphasized, butless involved in the active perception and learning process of visual nervous system. So the convolutional neural net-works with sparsely connection thought, self learning mechanism and good network topology structure are chosen asthe framework. Based on classical convolutional neural network model, with hierarchical and biomimetic idea, anew enhanced level convolution neural network（CNN） model IH-CNN based on visual nerve is proposed. Experi-mental results show that target recognition issue in large scale images can be better solved through IH-CNN modeland the target recognition accuracy rate is 84%.

关 键 词：生物视觉 目标识别 Caltech-101 卷积神经网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]