Restoration of ancient Chinese paintings using GANs with multi-scale defect fusion architecture and multi-pooling enhanced spatial-channel attention

Abstract

<title>Abstract</title> This study investigates the application of generative adversarial networks combined with multi-scale defect fusion architecture and multi-pooling enhanced channel–spatial in the restoration of ancient Chinese paintings. This approach addresses some challenges encountered during the restoration process, such as insufficient feature extraction and the restoration of minute textural details. To enhance the precision of models in restoring ancient paintings, we propose an enhanced generative adversarial network that integrates Fourier convolution. This model uses multi-scale defect fusion architecture (MDFA) to improve edge continuity in the repaired region by using parallel hollow convolutions and scenario fusion. Meanwhile, a multi-pooling enhanced channel–spatial (MECS) module uses robust attention mechanisms to enhance texture details. Through extensive experimentation on Chinese ancient painting datasets, this method significantly improves visual fidelity, brushstroke coherence and structural consistency compared to existing approaches. The results demonstrate that this model effectively restores the textural details of ancient paintings, which is superior to other models both qualitatively and quantitatively. In addition, it has been verified through the restoration of actual damaged ancient paintings, confirming its practicality and effectiveness.