ERF-FEM: The Feature Extraction Module based on Effective Receptive Field used for Image Inpainting

Many image inpainting models that rely on convolutional neural networks (CNN) have been extensively utilized in diverse image inpainting contexts, achieving outstanding results. Regrettably, due to the constraints of convolutional kernels, these approaches lack the capacity to capture multi-scale features. On the contrary, the majority of image inpainting methods regarded corrupted images as common pixel consistency images, disregarding the influence of the corrupted regions, such as shifts in mean and variance. To address the aforementioned challenges, we propose an image inpainting model that leverages a multi-scale feature extraction module and attention-enhanced learnable region normalization. It is capable of obtaining feature maps of different scales without stacking multiple layers of convolutions, alleviating the problems of detail information loss and low local computational efficiency. Additionally, during normalization, the learnable region normalization enhanced by attention can mitigate the offset effect of invalid pixels on mean and variance. Our method was verified on datasets Places2 and CelebA, and it was indicated by the experimental results that the similar methods currently accessible were outperformed by our method.