Face Image Restoration Research Based On Context Encoders And CGAN Joint Optimization

With the rapid development of science and technology,face recognition technology has also achieved remarkable development.In the process of face recognition,since there is generally no constraint condition when acquiring a face image,the posture,expression,and illumination of the person are complicated and variability,and sometimes there is an occlusion of decorations such as hats,glasses,and scarves.In the situation,these change factors will lead to the decline in the performance of existing face recognition technology,which seriously hampers the practical application of face recognition technology.In this case,it is necessary to repair the blocked face image.In order to be able to better recognize the face image,it is necessary to perform repair and restoration of the face image with occlusion or defect before recognition.Because the occluded or missing portion usually contains critical content(eg,eyes,mouth,etc.)that affect the entire facial change,it is necessary to fill in the missing region based on the existing undefective usable visual data.The existing image restoration methods for face defects mainly include partial differential equations,sparse coding,texture synthesis,and the like,among which the following problems exist:1.Most of the face image repair methods need to have the object to be repaired.The original complete sample is the premise.2.The patch search repair method is usually based on local or non-local information to repair the defect area of the image,and most existing methods only extract useful information from a single image or search patches from a known area of the same image.This results in content similar to the known region.This type of method is usually difficult to repair due to the lack of local and global information correlation and the large defect area.3.Based on repair methods such as sparse or deafness encoding,the repair result lacks fine texture details or produces visible artifacts around the boundaries of the defect area,and may even produce blurred or unrealistic images.In view of the inconsistency of the local and global information existing in the patch search and repair methods in the existing face defect image restoration methods and the problem of fuzzy images generated by the sparse coder,this paper proposes a Context Encoders and Conditional Generative Adversarial Nets combination Optimized model:Using the Context Encoders to understand the entire content image,the encoder obtains the corresponding high-dimensional features through mapping,and then decodes the high-dimensional features out of the missing region content based on the neural network through the decoder,providing semantic-based and missing parts.The reasonable prediction of the structure is to initially repair the face map.The face map will be initially repaired as a condition to generate input to the network,and further repairs may be performed to generate a patch of the large-area defective content without requiring the original face image of the object to be repaired.The consistency between the local content and the global content of the restored image is guaranteed and it is more clear and natural,In this paper,a wide range of experiments are performed on the CelebA face dataset,and the experimental results are qualitatively and quantitatively analyzed.In the qualitative assessment,the results of the model and the PM and CE models are compared in terms of visual perception;in the quantitative assessment,the peak signal-to-noise ratio(PSNR)and structural similarity index(SSIM)of the repair results are performed.A comparative analysis.The comparison results show that the repair results of this model are more realistic and clearer,and are more effective for large-area defect repair.