| The existing image inpainting methods mainly focus on images of natural scenes,buildings,etc.,and there has not been too much in-depth research on the inpainting of face images.The inpainting of face images under natural conditions will cause deviations in the inpainting results due to the facial posture,occlusion,expression and other factors,which is prone to the problems of fuzzy boundary area and incoherent structure.This topic focuses on the research on the inpainting of face images,and proposes an image inpainting algorithm based on the prior-guided face structure feature.In this algorithm,the potential information of face structure is extracted by using convolutional neural network in incomplete face images,and then the reconstructed image results are further improved through constraint of face structure conditions and reconstruction of hidden layer content.The main work of this paper are as follows:Firstly,in order to solve the problem of landmarks detection in incomplete face images,a detection network for landmarks in incomplete face images was constructed based on VGG16 and Mobile Net-V2 respectively,and a detection network for landmarks based on feature fusion was proposed.The feature extraction of the incomplete face image is carried out through the convolutional layer.Before the prediction of the mark points,the features of different layers are fused to realize the integration of effective information at different scales,so as to improve the detection accuracy of the mark points.Secondly,in order to improve the structural details of the face image inpainting result,the prior information of the face landmark structure is embedded in the face inpainting network.Compared with natural scenes,face images have stronger topological structure and attribute consistency.For the incomplete image to be repaired,the corresponding face landmark feature map is predicted in advance,and it is used as a priori information to guide the training of the image repair network.At the same time,in order to optimize the reconstruction of the network,add the expanded residual to extract deeper feature information At the same time,the resolution of the features is guaranteed to prepare for the subsequent reconstruction work.Finally,in order to optimize the image reconstruction sub-network and improve the quality of image restoration,a branch network based on deep supervision is proposed.By adding soft constraints to the features of the middle hidden layer of the reconstructed subnetwork,the prediction value is standardized from the features of the middle layer,and then the repair accuracy is gradually improved.At the same time,the joint up-sampling module is used to realize the feature fusion between the shallow features and the deep features,thereby ensuring the overall and detailed optimization of the repair results.The experimental model in this article is built and run based on the Pytorch environment,in which the image inpainting uses the open source Celeb A face image dataset,which contains 200,000 face images.In addition,in order to strengthen the predictive ability of the face landmark detection network,the common face landmark detection data sets 300 W and WFLW containing multiple categories are used to further generalize the detection of facial landmark points.The experimental results show that the landmark detection network based on feature fusion can solve the problem of incomplete face landmark detection.At the same time,the image reconstruction with the introduction of structural prior information can better repair the texture and structural details of the face image,and improve the face.Image inpainting effect. |
