International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 09 Issue: 09 | Sep 2022
p-ISSN: 2395-0072
www.irjet.net
Facial Image Restoration Using GAN Deep Learning Model Deepa M. Yadav1, Prof. Suraj S. Redekar2 1Student,
Dept. of Computer Science and Engineering, AMGOI College, Maharashtra, India Dept. of Computer Science and Engineering, AMGOI College, Maharashtra, India ---------------------------------------------------------------------***--------------------------------------------------------------------2Guide,
Abstract - Facial image priors, such as the facial geometry
outputs, but low accuracy images are produced as output. These methods not good at accurate restoration guidance.
prior or the facial reference prior, are required for facial image restoration. These are utilized to restore details and naturalness of facial feature. The extremely poor and significantly damaged inputs are unable to provide precise geometric prior. And it's difficult to find references of good quality. The facial prior's application in practical situations is constrained by these factors. The GAN is performing tremendously better in image generation. Therefore, this GAN may restore facial images by leveraging rich and diversified priors. The GAN is utilized during the process of restoring the face, because of its complex design as well as power use of generative facial prior. This GAN could perform both restore facial details as well as enhance quality and colors. This GAN may achieve superior performance on both real-world and synthetic datasets. Now a days more attention is taken by image generation. Many industrial applications looking for facial and text image restoration without losing its identity, that is the important thing to achieve. Generative adversarial networks are excellent in image editing and has the great potential to produce natural images.
To overcome these issues, the proposed GAN method may achieve more natural and accurate images. The degradation removal and facial features restoration are the important steps, which impacts the performance of the process.
2. LITERATURE REVIEW CNN models achieved great success in facial image restoration. It performs de-blurring, improving the resolution and many other face processing jobs with high accuracy. CNN analysis the images with its layers. In order to rebuild the high-resolution output from a very low-resolution facial image, Huang et al. suggested a CNN model that uses wavelet coefficients1. For recovering the photos, Cao et al. recommended a reinforcement learning technique2. The face hallucination approach uses a recurrent policy network to specify the subsequent attended region. The local enhancement network is then used to recover it. In order to reduce face image blurring,
Key Words: Image Restoration, Image Processing, GAN Model, Neural Networks, Deep learning
Chrysos et al. devised a technique that takes advantage of the well documented structure, description, and details of face3. And Xu et al. came up with a generative adversarial network for de-blurring of face and text4. The global semantic face priors can be used to restore the shape and details of face images. This technique is explored by Shen et al5. So, these are the existing single image restoration methods. The cons with these are, they perform poorly to real-world low quality, degraded face images. Also, they perform not so good due to the different poses and variety of severe degradations.
1. INTRODUCTION Facial image restoration has the goal of recovering the highquality images from the low-quality images. These lowquality images may have the degradation, which is consist of low resolution, noise, blurriness etc. It is very difficult to recover these images due to severe degradation, variety of poses, and different expressions. The work happened till now took the approach of using face specific priors to restore the facial image. These priors used to create the face shape and feature details. Another approach of using the of using the facial component dictionary. Which is used to generate the more realistic and natural images.
In some CNN approach images are enhanced by transferring intensity images’ structural details. One is guidance and another is degraded image, both are used in the face restoration process. According to Zhang et al. used a lengthy and difficult searching method by using a reference image with similar content6. In the space of features, it is applied to map high resolution image as a guidance with a lowresolution deteriorated patch.
So, here the deep learning method can leverage the capability of GAN that generates the facial images. This can generate more natural faithful faces with different variety. This creates the diverse and rich geometry priors, with variety of texture and colours. This enables the recovery of facial characteristics, and the improvement of colours Use the generative priors for real-world face restoration using pertained face Generative Adversarial Network (GAN) model. Previous attempts visually give the natural, soft
© 2022, IRJET
|
Impact Factor value: 7.529
Facial image restoration may be done using the predefined three-dimensional parameterized models or CNN, which is used to represent the face. This method has the capability to describe the faces, different poses, and deviated head positions. But this method is unable to describe the complex expressions and facial postures.
|
ISO 9001:2008 Certified Journal
|
Page 992