Research On Face Occlusion Generation Technology Based On Generative Adversarial Networks

Face recognition technology has developed significantly in recent years.However,identifying a partially occluded face is still a challenge to existing face recognition techniques.In practical applications,there is an increasing demand for occluded face image restoration,such as monitoring and public safety.In order to better recognize the face image and improve the recognition efficiency,it is necessary to repair and restore the occluded face image before recognition,because the occluded part usually contains key feature information(such as nose,eyes,etc.)that affects the entire face change.Mouth,etc.)The repair of face images is an important topic in the field of image restoration.Image restoration is a common image editing operation designed to fill missing areas in an image with reasonable content.The generated content needs to be similar to the original content,but also fully conforms to the overall image,making the restored image look real.In the past few decades,due to its inherent ambiguity and the complexity of natural images,image restoration(image restoration)has been a challenging research hotspot in the field of computer vision and image graphics.This paper combines the knowledge of traditional image restoration based on structure and texture,image restoration based on deep learning and generative adversarial networks,and analyzes the key points and structural feature information of face.Based on Boundary Equilibrium Generative Adversarial Networks(BEGAN)generate and repair the occluded face image.First train BEGAN on the CelebA dataset,then input the occluded face image into BEGAN,try to generate and repair the occlusion part;secondly,in order to make the generated image The original face image is more similar and more natural.Two discriminator networks are defined in the algorithm: the global structure discriminator network is responsible for optimizing the global edge structure information and feature information of the face repair image to ensure that the restored face image result conforms to visual cognition;The local texture discriminator network is responsible for optimizing the generated occlusion region to be similar to other facial content of the original face image(for example,skin texture,color,etc.),so that the generated face image is more natural and coherent.Finally,the loss function of the global and local two discriminator networks is integrated.The back-propagation algorithm is used to map the face image to be repaired to a smaller potential space,and the mapped vector data is input into BEGAN to generate the best face.Fix the image to achieve the generation and repair of the face occlusion image.In this paper,a large number of experiments are carried out in the CelebA face image dataset,and the experimental results are analyzed and compared,mainly subjective evaluation and objective evaluation.In the subjective evaluation,the main reason is to let the inside and outside of the line visually sense the face image restoration effect;in the objective evaluation,compare the model with the PatchMatch model and the Context Encoder model,using the peak signal to noise ratio.And the structural similarity index to compare the pros and cons of the model repair effect.The results show that the model can achieve good restoration of the occluded face image and is superior to the two models for comparative experiments.