Image Inpainting in Frequency Domain with Wavelet Convolution
Jain-Kai Huang, Tsung-Jung Liu, Kuan-Hsien Liu · 2022 IEEE International Conference on Consumer Electronics - Taiwan · 2022
This paper shows that Time-Frequency Analysis (TFA) techniques for signal processing can be used on tasks of computer vision. The idea is as follows: we can build a simple network architecture like convolutional neural network (CNN), analyze hidden features by wavelet transform or other methods, and catch them into filters for weights by convolutions, transformers or other ways. It looks like we need to build the network with 2 stages to accomplish the idea. However, we actually can directly use TFA skills in one-stage network by some technique. Networks which build from this way not only has good performance, but also cost lower computing resources. In this paper, we mainly use wavelet transform on CNN to solve free-form image inpainting problems. And it shows our CNN model can nicely work in frequency domain.