基于卷积神经网络(CNN)的VGG-19(Visual Geometry Group)模型,研究了卷积神经网络对输入纹理进行卷积时,输入纹理特征图的边缘信息对生成自然纹理效果的影响。在使用卷积神经网络的VGG对输入图像进行卷积时,为了防止过拟合现象,采用平均池化的方式对特征图进行处理,在一定程度上保护了特征图的边缘信息,相对采用最大池化处理特征图取得了更好的生成效果。同时,提取各层特征图的边缘信息并将其叠加到特征图中,能很好地保留纹理图像的边缘结构信息。实验结果表明,改进后的方法能取得较为理想的纹理生成效果。

Based on the Visual Geometry Group (VGG-19) model of convolutional neural networks (CNN), influences of the edge information in an input texture feature map on the natural texture are studied when the CNN convolves the input texture. When the input image is convoluted by the VGG using the CNN, the feature map is processed in an average pooling manner to prevent overfitting, which protects the edge information of the feature map to some extent and the generation effect is better than that obtained via max-pooling processing. The edge information of each layer of feature map is extracted and superimposed on the feature map, which preserves the edge structure information of the texture image well. Experimental results demonstrate that the proposed method achieves a good texture generation effect.