Efficient Perceptual Loss based on Low-rank Tucker Decomposition
Taehyeon Kim, Heungjun Choi, Yoonsik Choe · 2022
The perceptual loss function has been used successfully in image transformation for capturing high-level features from images in pre-trained networks. Standard perceptual losses require numerous parameters to represent features in networks; thus, it is not proper to resource-constrained devices. Hence, we propose a compressing perceptual-loss-oriented low-rank Tucker decomposition optimized with High-Order Orthogonal Iteration. Additionally, to decide an optimal low-rank in decomposition, we used variational Bayesian matrix factorization. In the technique thereof, salient features are extracted more efficiently. To the best of our knowledge, we are the first to consider curtailing redundancies in feature maps via low-rank Tucker decomposition. Experimental results in style transfer tasks demonstrate that our method not only yields similar qualitative and quantitative results as that of the original version but also reduces memory requirement by approximately 34%.