摘要
卷积神经网络中的池化操作可以实现图像变换的缩放不变性,并且对噪声和杂波有很好的鲁棒性。针对图像识别中池化操作提取局部特征时忽略了隐藏在图像中的能量信息的问题,根据图像的能量与矩阵的奇异值之间的关系,并且考虑到图像信息的主要能量集中于奇异值中数值较大的几个,提出一种矩阵2-范数池化方法。首先将前一卷积层特征图划分为若干个互不重叠的子块图像,然后分别计算子块图像矩阵的奇异值,将最大奇异值作为每个池化区域的统计结果。利用5种不同的池化方法在Cohn-Kanade、Caltech-101、MNIST和CIFAR-10数据集上进行了大量实验,实验结果表明,相比较于其他方法,该方法具有更好地识别效果和稳健性。
The pooling operation in convolutional neural networks can achieve the scale invariance of image transformations, and has better robustness to noise and clutter. In view of the problem that pooling operation ignores the energy information hidden in the image when it extracts local features for image recognition, according to the relationship between energy of the image and singular value of the matrix, and taking into account the image information of the energy mainly concentrates on the larger singular value, a pooling method based on matrix 2-norm was proposed. The former feature map of convolutional layer is divided into several non-overlapping sub blocks, and then singular value of the matrix is calculated. The maximum value is used as the statistical results of each pooling region. Various numerical experiments has been carried out based on Cohn-Kanade, Caltech-101, MNIST and CIFAR-10 database using different kinds of pooling method. Experimental results show that the proposed method is superior in both recognition rate and robustness compared with other methods.
出处
《图学学报》
CSCD
北大核心
2016年第5期694-701,共8页
Journal of Graphics
关键词
深度学习
卷积神经网络
矩阵2-范数
池化
奇异值
deep learning
convolutional neural networks
matrix 2-norm
pooling
singular value