DWTSISR: a discrete wavelet-based approach to lightweight super-resolution reconstruction

To address the issues of low contrast, blurred details, and blurred edges in images, a super-resolution reconstruction network DWTSISR that integrates discrete wavelet transform (DWT) is proposed and demonstrated in this paper. DWTSISR can effectively preserve key information of images and reduce data volume through multi-scale decomposition of DWT, thereby reducing processing complexity. In particular, the integration of DWT with physical optics priors enables the method to not only efficiently extract features in the wavelet domain but also reconstruct high-quality images in the spatial domain, overcoming the computational bottlenecks present in traditional methods. First, the network performs DWT decomposition on the input image to generate sub-band coefficients at multiple scales. Then, these sub-band coefficients are processed in the wavelet domain using a lightweight network architecture that includes multiple convolutional layers and activation functions to extract and enhance features. The processed sub-band coefficients are reconstructed back into the spatial domain through inverse DWT, generating high-quality images. To further optimize the model’s performance in various complex scenarios, a novel hybrid loss function, to our knowledge, is introduced, which combines image reconstruction loss and perceptual loss to improve the details and visual quality of the reconstructed images. The super-resolution reconstruction network DWTSISR was experimentally analyzed on multiple datasets, and the experimental results show that DWTSISR exhibits excellent performance on several benchmark datasets, maintaining high levels of PSNR and SSIM values. On the CUMID mine dataset, the PSNR value of DWTSISR is 0.3 higher than that of the optimal algorithm RCAN, while the parameter count is only 0.5.