Multi-scale adversarial generative network for super-resolution of endoscopic images

High-definition endoscopic images can provide clinicians with accurate marginal information and complex lesion texture details. However, in endoscopic surgery, the image resolution is reduced due to the limited resolution of the acquisition camera, the imaging noise introduced by the endoscopic light source, and the physiological movement that causes motion artifacts in the texture of the lesion edge. To address these issues, a super-resolution reconstruction method based on conditional adversarial generative networks is proposed. To tackle the problem of blurred edge textures, a multi-scale mixed attention feature extraction module is designed in the down sampling stage, which extracts image features and enhances the weight of local areas. In the up sampling stage, high-frequency details extracted from the Laplacian pyramid are introduced. To effectively remove light source noise and improve the reconstruction effect of edge textures, a novel hybrid perceptual function is proposed that fuses content loss, style loss, total variation loss, perceptual loss, and texture loss based on L1 loss and adversarial generation loss. The discriminator uses a relative discriminator to enhance the quality of the generated images by judging how much more realistic the real images are compared to the generated ones. The reconstruction results on the Kvasir dataset and a private dataset show that the PSNR index increases by an average of 5.34 dB, the SSIM index increases by 0.10 on average, the PIQM decreases by 0.015 on average, and the LPIPS decreases by 0.074 on average, indicating that the proposed method demonstrates significant advantages in detail representation and lesion texture reconstruction, and can effectively suppress noise to improve overall visual quality.