Complementary Transformer Network for cross-scale single image denoising

Cliffside carving images are often affected by various types of noise, such as uneven lighting, shadows, dust, and weathering, which impair the clarity and detail of the images. These noise factors significantly impact image quality, making effective denoising crucial. Denoising can enhance the clarity and quality of cliffside carving images, facilitating the study of their artistic style, historical background, and cultural significance. Therefore, this paper proposes the Complementary Transformer Network (CoTrNet), which utilizes an encoding-decoding framework to denoise cliffside carving images. The Diverse Feature Complementary Module (DFCM) is employed for feature extraction and image reconstruction, while the Skip Connection Cross Transformer (SCCT) enhances the transfer of low-level features to higher levels, improving the overall denoising effect. CoTrNet accurately captures the details and features in the images, significantly reducing noise. Using images from Tongtian Rock in Ganzhou, China, experiments show that CoTrNet outperforms existing techniques, achieving Peak Signal-to-Noise Ratios (PSNR) of 27.5706 and 24.9113 at noise levels of 15% and 25%, respectively. This research provides a powerful tool for the preservation, restoration, and conservation of cliffside carving cultural heritage. • We introduce the Complementary Transformer Network (CoTrNet) for denoising cliffside carving images, ensuring global consistency and capturing subtle details. The CoTrNet significantly reduces noise and enhances the clarity of the cliffside carving images.