Rethinking and Improving the Robustness of Image Style Transfer
This work addresses robustness issues in image style transfer for AI and computer vision applications, offering an incremental improvement by modifying feature extraction.
The paper tackled the problem of neural style transfer's poor performance with advanced networks like ResNet, finding that residual connections reduce feature entropy and proposing a softmax transformation to enhance it, which improved stylization quality even with random weights.
Extensive research in neural style transfer methods has shown that the correlation between features extracted by a pre-trained VGG network has a remarkable ability to capture the visual style of an image. Surprisingly, however, this stylization quality is not robust and often degrades significantly when applied to features from more advanced and lightweight networks, such as those in the ResNet family. By performing extensive experiments with different network architectures, we find that residual connections, which represent the main architectural difference between VGG and ResNet, produce feature maps of small entropy, which are not suitable for style transfer. To improve the robustness of the ResNet architecture, we then propose a simple yet effective solution based on a softmax transformation of the feature activations that enhances their entropy. Experimental results demonstrate that this small magic can greatly improve the quality of stylization results, even for networks with random weights. This suggests that the architecture used for feature extraction is more important than the use of learned weights for the task of style transfer.