Research On Face Image Inpainting Method Based On Multi-scale Convolutional Fusion

As a carrier of face information,face images play an irreplaceable role in various research fields.However,when the environmental conditions of image capture are poor,the captured face images may have different degrees of obscuration,or perhaps due to improper processing in the process of image transmission,storage and post-processing,resulting in the inability to obtain a complete and clear face image.Obscured or blurred face images not only lead to poor visual perception,but also may cause serious security risks and unpredictable economic losses in some important situations.Therefore,it is very meaningful to study how to repair the obscured or blurred face images into real images.In recent years,with the rapid development of artificial intelligence technology,many researchers have done a lot of work on face image inpainting tasks and have achieved different degrees of results.However,there are still two thorny problems in the field of face image inpainting.The first problem is the poor restoration of face images with large missing regions,and the second problem is that the current research methods cannot effectively use face a priori information to assist image restoration.To address the above two problems,the main research of this thesis is as follows.1.In order to improve face image inpainting with large missing regions,this thesis designs an attention mechanism-based multiscale convolutional fusion module(MFMAM),which consists of a hybrid attention mechanism and parallel dilated convolution to better extract the rich remote contextual information in images.2.In order to make better use of face priori information in the face image inpainting process,this thesis proposes a novel face image inpainting network(SS-MFM-AM-GAN)based on face priori information guidance,which consists of three parts,the first part is an MFM-AM-GAN incorporating moment shortcut with positional normalization(PONO-MS)for rough restoration of face images;the second part is a face semantic segmentation network that for extracting semantic information from the rough restoration results of the first part;the third part is a decoder designed based on spatially-adaptive(DE)normalization(SPADE)for incorporating the semantic information of faces from the second part into the image restoration task.In this thesis,the above method is compared with several classical face image inpainting methods on the datasets CelebA,CelebA-HQ and EDFace-Celeb-1M,and the experimental results are analyzed using objective evaluation metrics and subjective visual perception,and the results show that the attention mechanism-based multiscale convolutional fusion module designed in this thesis has improved the restoration effect for large-area missing images,while the face priori information-based face image inpainting network proposed in this thesis can better utilize the face priori information to assist the face image restoration work.In addition to the comparison experiments,ablation experiments are also designed to demonstrate the effectiveness of the above research work.