Computationally Efficient Approaches for Image Style Transfer
Ram Krishna Pandey, Samarjit Karmakar, A. G. Ramakrishnan · 2018
In this work, our focus is on developing fast image style transfer architectures for practical applications. We have proposed three modifications to the architecture of a recent, real-time, artistic style transfer technique to make it computationally more efficient. We have proposed the use of depth-wise separable convolution (DepSep) in place of convolution and nearest neighbor (NN) interpolation in place of transposed convolution. We have also explored the concatenation of nearest neighbour and bilinear (Bil) interpolations in place of transposed convolution. The stylized images from the modified architectures are perceptually similar in quality to those from the original architecture. The decrease in the computational complexity of our architectures is validated by the decrease in the testing time by 26.1%, 39.1%, and 57.1%, respectively, for DepSep, DepSep-NN-Bil and DepSep-NN modifications. Working with another architecture, we have examined how the quality of the stylized reconstruction changes with the change of the loss function to be minimized.