摘要
针对传统卷积神经网络CNN(Convolutional Neural Networks)在训练或学习时只利用图像的灰度信息,丢失了颜色信息的问题,提出一种基于多通道卷积神经网络来提取特征的方法。该算法对于每一个颜色通道分别学习一个多层卷积神经网络,并且在输出层通过全连通的神经网络进行融合。算法首先建立三个多层卷积神经网络来学习图像三个通道(RGB,HSV,Lab等)的特征;然后将三个颜色通道的特征赋予不同的权值(权值和为1)后进行融合,得到样本的特征;最后通过一个全连通的神经网络得到分类结果。实验结果分析表明,该算法相比于传统卷积神经网络能取得更高的准确性,同时能更好地适应复杂多变的环境。
Conventional convolution neural network only uses image gray scale information in training or learning but loses colour information. Aiming at this problem, the paper proposes a method which extracts the features based on multi-channel convolution neural networks. The algorithm learns one multi-layers convolution neural network for each colour channel respectively and fuses them through a full connected network on output layer. First, the algorithm establishes three multi-layers convolution neural networks to learn the characteristics of three channels （ RGB, HSV, Lab, etc. ） of image. Then it assigns different weights （weights sum 1 ） to the characteristics of these three colour channels and then fuses them to get the sample characteristics. Finally, it obtains the classification results by a fully connected neural network. It is showed by analysing the experimental results that this algorithm has higher accuracy than conventional convolution neural network while can better adapt to complex changing environments.
出处
《计算机应用与软件》
CSCD
2016年第1期159-162,共4页
Computer Applications and Software
关键词
卷积神经网络
颜色信息
多通道
自学习特征
Convolution neural network Colour information Multi-channel Self-learning feature
