An image inpainting model based on channel attention gated convolution and multi-level attention mechanism

Abstract

Deep learning-based image inpainting models can effectively improve the clarity and restoration of damaged images. Most inpainting models fail to fully leverage deep semantic information, which leads to structural inconsistencies or missing details, especially when dealing with images containing large regions of missing information. This paper introduces an image inpainting model that integrates channel attention gated convolution (CAGC) and multi-level attention mechanism (MAM), referred to as CAGC-MAM model. The model is based on a generative adversarial network (GAN) and employs a two-stage coarse–fine network structure. Through the CAGC module, it effectively captures global features, mitigating the limitations of standard convolutional receptive fields. The MAM module combines spatial and channel attention to precisely extract critical information, ensuring the accuracy of restored details. Moreover, a joint loss function is applied to enhance detail consistency further. Qualitative and quantitative analyses conducted on CelebA-HQ and Paris Street View datasets, indicate that CAGC-MAM model excels in image clarity, detail enhancement, and overall consistency, showing superior performance in both computational time and inpainting effectiveness compared to existing classic models.