Conditional Generative ConvNets for Exemplar-Based Texture Synthesis.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Exemplar-based texture synthesis aims to create visually similar textures from a given example.
• Current methods using pre-trained Convolutional Neural Networks (ConvNets) struggle with non-local structures and dynamic/sound textures.

Purpose of the Study:

• To develop a unified model for synthesizing diverse texture types (image, dynamic, sound).
• To overcome limitations of existing deep learning models in texture synthesis.

Main Methods:

• Introduced a conditional generative ConvNet (cgCNN) model combining deep statistics and a generative ConvNet (gCNN) framework.
• cgCNN learns ConvNet weights per exemplar, unlike models relying on pre-trained networks.
• Synthesizes textures by sampling from a conditional distribution defined by deep statistics.

Main Results:

• Achieved high-quality synthesis of dynamic, sound, and image textures in a unified manner.
• Demonstrated superior or comparable performance against state-of-the-art methods.
• Showcased generalization capabilities for texture expansion and inpainting.

Conclusions:

• The proposed cgCNN model offers a versatile and effective approach to texture synthesis.
• The model's ability to learn exemplar-specific weights enhances its performance on complex textures.
• cgCNN provides a unified framework for various texture synthesis tasks, including expansion and inpainting.