Image Inpainting Transformer with Mask-Aware Encoding and Dilated Convolutional Attention

Existing image inpainting methods usually use convolution-based downsampling to reduce the spatial dimension of the image, but this method may cause information loss when dealing with large missing areas. Due to the limitation of computational cost, the self-attention mechanism in Transformer cannot fully utilize the long-term modeling capabilities of these models. To solve the above problems, this paper proposes an innovative image inpainting model: Image Inpainting Transformer with Mask-aware Encoding and Dilated Convolutional Attention (IMD). The model uses Mask-aware Depthwise Separable Downsampling Module (MDD) to preserve the visible features of the damaged image. At the same time, the Dilated Convolutional Attention Layer (DCAL) is introduced to enhance the model's ability to capture image spatial and channel features by combining the self-attention mechanism and Dilated Convolutional, achieving multi-scale image inpainting effects. Experimental results on the Paris Street View standard dataset show that the IMD algorithm is significantly better than the existing technology. We conducted experiments on the Dunhuang mural dataset, and the experimental results also verified the feasibility of this method in cultural relics inpainting.