基于卷积神经网络与Bayesian决策的图像识别与分类记忆建模

发布时间：2018-03-04 19:47

本文选题：卷积神经网络　切入点：Bayesian决策规则　出处：《中国科学:技术科学》2017年09期 　论文类型：期刊论文

【摘要】：当前热门图像分类方法大多侧重在分类能力,忽视识别新事物,然而人类认识事物时侧重认识,只在细小之处重视分类,这一点与人类记忆机制密切相关.尽管目前有许多记忆建模理论被相继提出,但大多以单词列表的形式学习,对自然图像列表的研究有限.基于此,本文提出了基于卷积神经网络与Bayesian决策的图像识别分类记忆建模方法,首先利用卷积神经网络提取图像特征,并采用二进制形式存储特征向量;然后进行视觉图像的表达,存储与提取记忆建模,将测试图像特征向量与所有已存储特征向量进行匹配对比,计算似然率值;最后在所有似然率基础上计算测试图像是新类别的几率,若该几率大于某个阈值则判别其为新类别;反之,利用Bayesian决策规则进行.图像分类.在Caltech-101与Caltech-256数据库上的实验表明所提方法能很好地应用于图像识别分类任务中.其击中率比目前代表性的稀疏表达分类(SRC)以及极限学习机(ELM)方法高,且虚报率比其他两种方法低的多.
[Abstract]:At present, most popular image classification methods focus on the ability of classification, ignoring the recognition of new things. However, when people know things, they pay attention to classification only in small places. This is closely related to the mechanism of human memory. Although many theories of memory modeling have been put forward one after another, most of them are studied in the form of word lists, and the study of natural image lists is limited. In this paper, a method of classifying memory modeling based on convolution neural network and Bayesian decision is proposed. Firstly, the image feature is extracted by convolution neural network, and the feature vector is stored in binary form, then the visual image is represented. The model of storing and extracting memory is used to match the feature vector of test image with all the stored feature vectors, and the likelihood rate is calculated. Finally, the probability that the test image is a new category is calculated on the basis of all likelihood rates. If the probability is greater than a threshold, it is identified as a new category; conversely, The experiments on Caltech-101 and Caltech-256 database show that the proposed method can be well applied to the task of image recognition and classification. The hit rate of the proposed method is better than that of the representative sparse representation classification and pole. Learning-limited machine (ELM) method is high, And the false report rate is much lower than the other two methods.
【作者单位】：中国石油大学(华东)信息与控制工程学院;
【基金】：国家自然科学基金项目(批准号:61271407,61671480) 中央高校基本科研业务费专项资金项目(编号:14CX06066A)资助
【分类号】：TP183;TP391.41

【相似文献】