Fresco Arbitrary Scale Super Resolution Reconstruction Based on Frequency Domain Transform
Jiaqi Liu, Jianfang Cao · 2024
In response to the problem that traditional super-resolution models with low reconstruction magnification cannot perfectly reconstruct the details of high-scale murals and cannot present the texture elements of the original painting, the paper proposes a fresco arbitrary scale super-resolution reconstruction based on frequency domain transform. Firsty, the paper fuses the image features of the mural with spatial coordinates. It builds a frequency domain transformation module based on the encoding-decoding architecture to transform the mural image in the frequency domain, learning the frequency, amplitude, and phase characteristics of the mural and enhancing the implicit feature expression of the texture. Secondly, employs a multilayer perceptron as a decoder to learn implicit feature representations of images, thus achieving arbitrary-scale super-resolution. Lastly, constructs a Siamese neural network and uses the trained model as a reconstruction constraint condition for the leading network. Experimental results show that on the mural dataset, compared with popular arbitrary-scale super-resolution reconstruction models, the experimental results show that the PSNR and SSIM of this method are improved by at least 0.09 dB and 0.007 respectively. This method has a better performance in reconstruction quality and image structure.