Compact Image-Style Transfer: Channel Pruning on the Single Training of a Network.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Current image-style transfer methods rely on VGG networks, which are computationally intensive and memory-demanding due to their design for general image classification.
• The redundancy in VGG networks is unnecessary for the specific task of image style transfer, leading to inefficiencies.

Purpose of the Study:

• To propose a novel technique for downsizing style transfer networks by eliminating redundant channels in VGG feature networks.
• To reduce memory consumption and computational cost associated with image style transfer.

Main Methods:

• A new method is introduced to automatically identify and deactivate convolution channels during network training.
• Two novel losses, 'channel loss' and 'xor loss', are employed to maximize inactive channels and fix their positions, respectively.
• These losses are also applied to prune the VGG16 classifier network.

Main Results:

• The proposed method accelerates image generation speed by up to 49%.
• It reduces the number of parameters by 20% in style transfer networks while maintaining performance.
• For VGG16 classifier networks, parameter reduction reached 26% with a 0.16% improvement in top-1 accuracy on CIFAR-10.

Conclusions:

• The developed technique effectively reduces the computational and memory footprint of image style transfer networks.
• The method offers significant speedups and parameter reduction without compromising the quality of style transfer.
• The introduced losses demonstrate versatility by also improving the efficiency and accuracy of image classification networks.