应用卷积神经网络的纹理合成优化方法

Texture synthesis optimization method based on convolutional neural network

针对传统纹理合成方法特征提取困难以及合成周期较长的问题,提出一种基于卷积神经网络的纹理合成优化方法。通过优化VGGNet卷积神经网络的结构,并提出增加批量归一化BN层的方法,来提高网络训练速度和减少参数过拟合现象;通过计算每层得到的纹理图像特征响应的克莱姆矩阵,构建克莱姆矩阵集合来表达纹理特征;由梯度下降算法计算梯度,通过L-BFGS优化算法最小化损失函数,合成纹理图像。实验结果表明,该方法可以有效提高模型训练速度,减少参数过拟合现象,合成高质量的图像。

Traditional texture synthesis methods have many problems,such as hard feature extraction and the long period of texture synthesis.To this end,a texture synthesis optimization method based on the convolutional neural network was proposed.The network training speed was improved and the parameter over-fitting phenomenon was reduced by optimizing the structure of VGGNet and increasing the batch normalization layers.Gram matrices were built by obtaining each layer’s feature responses of texture image,and texture features were expressed by the set of Gram matrices.The gradients were computed using the gradient descent algorithm and the loss-function was minimized using L-BFGS optimization algorithm.The texture image was synthesized.The experimental results show that the proposed method can effectively improve the training speed of the network,reduce the parameter over-fitting phenomenon and synthesize high quality images.