Uncorrelated feature encoding for faster image style transfer

Recent image style transfer methods use a pre-trained convolutional neural network as their feature encoder. However, the is not optimal for but rather classification. Furthermore, they require time-consuming alignment to consider existing correlation among channels of encoded map. In this paper, we propose an end-to-end learning method that optimizes both encoder and decoder networks task relieves computational complexity correlation-aware alignment. First, performed updates only also parameters in training phase. Second, addition previous content losses, uncorrelation loss, i.e., total coefficient responses channels. Our loss allows generate map without correlation. Subsequently, our results faster forward processing with light-weighted transformer correlation-unaware Moreover, drastically reduced channel redundancy during process. This provides us possibility perform elimination negligible degradation generated quality. applicable multiple scaled by using cascade scheme user control strength through usage content-style trade-off parameter.