Blind face inpainting is the task of automatically recovering an occluded face image without given masks indicating missing areas. Popular inpainting methods assume that the occlusion patterns are known with given occlusion masks. Previous blind inpainting methods, ignoring the structure in faces and occlusions, treat occlusion detection as an independent pixel prediction problem. To overcome the limitations, we propose an enhanced two-stage blind face inpainting framework which consists of a structured mask prediction module and an inpainting module. The structured mask prediction module is first trained with a consensus loss for non-sparse and structured prediction. Then the inpainting module reconstructs the predicted missing areas and generates a visually plausible face image with a context normalization that enhances the robustness against prediction errors. We conducted experimental evaluations on the FFHQ and LFW datasets. The results demonstrate that our method is effective in producing visually convincing results with more continuous occlusion mask predictions and outperforms state-of-the-art methods in synthesized occluded face inpainting. Additionally, the method can effectively remove certain natural occlusions.

中文翻译：

---

通过结构化掩模预测增强盲人脸部修复

盲脸修复是在没有给定指示缺失区域的蒙版的情况下自动恢复被遮挡的面部图像的任务。流行的修复方法假设使用给定的遮挡掩模已知遮挡图案。以前的盲修复方法忽略了人脸和遮挡的结构，将遮挡检测视为独立的像素预测问题。为了克服这些限制，我们提出了一种增强的两阶段盲人脸部修复框架，该框架由结构化掩模预测模块和修复模块组成。结构化掩模预测模块首先使用非稀疏和结构化预测的共识损失进行训练。然后，修复模块重建预测的缺失区域，并生成视觉上合理的面部图像，并通过上下文归一化增强针对预测错误的鲁棒性。我们对 FFHQ 和 LFW 数据集进行了实验评估。结果表明，我们的方法可以通过更连续的遮挡掩模预测有效地产生视觉上令人信服的结果，并且在合成遮挡面部修复方面优于最先进的方法。此外，该方法可以有效地去除某些自然遮挡。