Fig. 2: Architectural comparison of the standard residual block and the enhanced CAAT block in the CAUGAN method.

a Standard Residual Block, b Improved CAAT Block. Note: This figure presents an architectural comparison between the Standard Residual Block and the Improved CAAT Block used in the CAUGAN method. In (a), the standard residual block processes the input feature \({x}_{1}\) through a 3 × 3 convolution and ReLU activation, then adds the result to the residual input \({x}_{2}\). While effective for general feature learning, this design lacks explicit spatial modeling, which can result in inconsistencies, such as color or texture mismatches in mural restoration. In contrast, b shows the CAAT Block, which introduces coordinate attention mechanisms and a refined structure to enhance spatial awareness. The convolution operation is divided into four sub-kernels with reduced output channels (e.g., 64 channels per sub-kernel from an original 256-channel kernel). Each sub-kernel applies a distinct dilation rate (1, 2, 4, and 8), expanding the receptive field to capture both fine details and broader contextual information. This multi-scale approach is essential for handling large or repetitive patterns often found in mural backgrounds. The outputs from these sub-kernels are concatenated and fused through an additional convolution, enabling the CAAT Block to integrate multi-scale features effectively and model spatial hierarchies. This enhanced spatial awareness helps preserve mural structure and detail, resulting in improved performance during inpainting tasks. Overall, the CAAT block outperforms the standard residual block by incorporating spatially-aware, multi-scale contextual learning, making it highly suitable for complex tasks such as mural image restoration.