W-EFLN: A Wavelet Transform-Based Extendable Frequency Learning Network for Image Super-Resolution

With the advancement of the economy and technology, visual information is exerting an unprecedented amount of influence on all facets of modern culture and life. Deep neural networks have been incredibly successful at achieving single image super-resolution. Most of these methods utilize features from the low-resolution image space exclusively, and ignoring the frequency relationship of the image, resulting in their limited ability to enhance the high-frequency details. To address this problem, we propose a wavelet transform-based extendable frequency learning network, which enables the reconstructed image to contain abundant detail information. By learning the multi-level wavelet coefficients of the target image, we can learn the high-frequency information of the images. The extendable branch of the method is formed by stacking residual distillation blocks which are composed of a high-frequency attention module and an improved deformable convolution using side convolution kernel. In order to improve reconstruction outcomes, we also train our method with loss constraints in the wavelet and spatial domains. Experimental results demonstrate that our proposed method achieves superior performance against the state-of-the-art super-resolution methods in terms of both subjective visual quality and objective quantitative metrics.