多通道卷积神经网络图像识别方法  

Image Recognition Method of Multi-channel Convolutional Neural Network

作　　者：易超人 邓燕妮[1] 

机构地区：[1]武汉理工大学自动化学院,湖北武汉430070

出　　处：《河南科技大学学报（自然科学版）》2017年第3期41-44,共4页Journal of Henan University of Science And Technology:Natural Science

基　　金：国家"863"计划基金项目(2015AA015904)

摘　　要：为了更好地利用图像数据中隐含的特征信息,将多方向梯度信息作为边缘信息的基本表达,提出了一种基于图像梯度的多通道卷积神经网络图像识别方法。先将图像进行Sobel算子处理,得到水平方向、垂直方向及两个对角方向的4个梯度图像。然后,建立4个多层卷积神经网络,学习4个不同方向梯度图像的特征。再将4个不同方向的特征进行随机化特征融合,得到样本的特征后经过批标准化处理。最后,通过分类器得到分类结果。在数据库Cifar-10和MNIST上进行了验证,验证结果表明:本文提出的模型具有较好的泛化能力,相比单通道卷积神经网络,在两个数据库中识别错误率分别降低了9.85%和0.38%。To make better use of the feature information implied in the image database,a multi-channel convolutional neural network image recognition method based on image gradients was proposed by making the multi-directional gradients information as basic expression of edge information. Firstly,four gradient images of horizontal,vertical,and two diagonal directions were obtained by processing the image through Sobel operation.Then,four multi-channel convolutional neural network was built to learn the features of gradient images of four different directions,and the features of four different directions were fused by stochastic feature fusion. Finally,the classification results were obtained through the classifier after batch normalization process for the features of samples. The proposed model was verified on database Cifar-10 and MNIST. The results show that the proposed model has better generalization ability,and its recognition error rates are reduced by 9. 85% and 0. 38%respectively compared to single-channel convolution neural network in the two databases.

关 键 词：卷积神经网络 多通道 梯度图像 随机化特征融合 分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]