Unpaired image-to-image translation with improved two-dimensional feature

With the feature-level constraints, unpaired image translation is challenging in generating poor realistic images, which focuses on convolutional feature extraction, ignoring the SVD feature extraction. To address this limitation, the Unpaired Image-to-image Translation with Improved Two-dimensional Feature (UNTF) is proposed. Specifically, in our method the novel feature extraction module consists two part: the SVD feature extraction and the convolutional feature extraction. The SVD feature maps were built by Two-Dimensional Feature which transform 1-D features into 2-D features to cascade with convolutional features. In up-sampling module sub-pixel convolution is used to replace transposed convolution. What’s more, the proposed feature loss can stabilize the training process of generator. Finally, the proposed network was verified by ablation study and state-of-the-art methods. Experiments on image translation, image illustration, and image restoration show that both the image clarity index (EGF) and experts agree that the proposed method is superior to the existing methods.