Texture Synthesis Using Improved Transfer Learning

Texture synthesis has been studied for more than two decades. Recent progress in deep learning is a great opportunity to revisit the texture synthesis based on convolutional neural networks. However, texture synthesizing is still a problem of trade-off generality for efficiency. In this paper different deep pre-trained Convolutional Neural Networks, usually used in classification problems have been considered. To our knowledge, a part of the VGG-19 model, none of the studied networks have been used in texture synthesis. Alexnet, GoogLenet, VGG16, VGG19 and ResNet 50 are used. Carried tests have shown that textures for various scales and structures have been well generated, surpassing results obtained by state-of-the-art methods.