Welcome to click on the top of the blue word, attention ~

Related reading: \

GAN is 6 years old! It’s time to stroke him!

Losing face! GAN to predict? \

Weak water three thousand, only take your mark! How about AL (Active Learning) and GAN? \

Anomaly detection, GAN how GAN?

Virtual change of clothes! Quick overview of these latest papers how to do! \

Facial makeup migration! Take a quick look at a few papers using GAN

[1] How about the generation of GAN in medical images?

01-GAN formula concise principle of ironclad small treasure chapter \

On this day, you walk down the street, watching the traffic and pedestrians.

Suddenly, a startling profile appears 15 meters ahead of you to your left, and you can’t help feeling startled:

“Such side yan, must be a flourishing era unparalleled, the world rare few back view?”

So you rush to find, to see the face of the world…

A few minutes later, frustrated, you complain:

“What a face! What a profile killer!”

Figure from the network, intrusion deletion.

Lost you, now can only open peppa Pig animation, a solution to sorrow.

“The pig always faces the audience with one side, but no one has ever seen the other. The strange thing is, even from the side, there are still two eyes and two nostrils.”

Then you touch your heart and ask your soul:

“Given a profile, how do you get a true face??”

Ha, ha, ha, ha! Today I am sorting out a paper on “face normalization” with GAN (wait to see). Given an image of a side face, how do you get a front face image?

1 (2020-1-6) PI-GAN: LEARNING POSE INDEPENDENT REPRESENTATIONS FOR MULTIPLE PROFILE FACE SYNTHESIS

Arxiv. Xilesou. Top/PDF / 2001.00…

We often hope that through some method to extract a face posture image fixed “posture representation feature”, and in order to generate a variety of view of the posture image, such as face (given a side face, you can deduce the face image). It is still a difficult problem how to obtain “gesture representation features” from a facial orientation image to synthesize other facial gestures. Face transfer has application value in various fields such as multimedia security, computer vision and robotics.

In order to solve this problem, this paper proposes PIGAN (Circular shared encoder/decoder framework), which uses generative adversance network (GAN) with encoder-decoder structure and joint discriminator network to learn and extract “postage-independent features”, and then realizes realistic face synthesis.

Compared with traditional GAN, it also consists of an auxiliary encoder-decoder frame, which shares weights with the main frame and reconstructs images from the original posture image. The main framework focuses on creating “decoupled representations”, while the secondary framework aims to restore the original face. The method was validated using a CFP high resolution dataset.

Experiment: CFP data set. Made up of 500 people, each containing 10 frontal images and 4 profiles. The model was trained on 450 people and evaluated for the rest. The basic framework architecture follows the DCGAN implementation. Here are the renderings, some of which look good at first glance?

2 (2019-02-26) BoostGAN for Occlusive Profile Face Frontalization and Recognition

Arxiv. Xilesou. Top/PDF / 1902.09…

There are many factors that can affect the face recognition effect, such as posture, occlusion, lighting, age and so on, among which the main problem is large posture and occlusion, which may even decrease the performance of the model by more than 10%.

Postural invariant feature representation and face turning using generative adversarial networks (GAN) have been widely used to solve postural problems. However, the recognition of occluded profiles is still a problem to be solved. Therefore, this paper provides an effective solution, even in the face of the key areas of the face (such as eyes, nose, etc.) damaged or blocked side face image, also try to recognize.

Specifically, a BoostGAN is proposed for occlusion removal, frontal recognition and face recognition. Based on the assumption that facial occlusion is partial and incomplete, images of multiple occlusion blocks will serve as inputs, known as knowledge Boosting, such as identity and texture information. Then, a new aggregation network module is further designed for the final fine image synthesis.

3 (2018-10-6) Learning a High Fidelity Pose Invariant Model for High-resolution Face Frontalization

Arxiv. Xilesou. Top/PDF / 1806.08…

Face positivity refers to the process of synthesizing the front view of the face according to the given side face. Due to occlusion and distortion, it is extremely difficult to recover good results and preserve texture details at high resolution. In this paper, a high-fidelity postural invariance model (HF-PIM) is proposed to produce realistic and consistent identity-preserving results.

4 (2018-3-4) Load Balanced GANs for Multi-view Face Image Synthesis

Arxiv. Xilesou. Top/PDF / 1802.07…

From a single image to synthesize multi-view face is an ill-posed morbid, uncertain problem, the result often has serious appearance distortion. Generating realistic, identity-preserving multiple views remains a challenge. In this paper, a load-balanced generative adversal network (LB-GAN) is proposed, which can accurately rotate the yaw Angle of the input face image to any specified Angle.

LBGAN breaks down the challenging synthesis problem into two sub-tasks: face standardization and face editing. Normalization starts by positising the input image, and then the editor rotates the positised image to the desired pose. In order to generate realistic local details, the normalizer and editor were trained in two stages and constrained by conditional self-cycle loss and L2 Los-based attention.

5 (2017-12-13) UV-GAN Adversarial Facial UV Map Completion for Pose-invariant Face Recognition

Arxiv. Xilesou. Top/PDF / 1712.04…

A recently proposed robust 3D face alignment method establishes a dense or sparse correspondence between 3D face models and 2D face images. The use of these methods is both a challenge and an opportunity for facial texture analysis. In particular, facial UV can be created by sampling the image using the FITTED model. D But the UV image is always incomplete due to occlusion. In this paper, UV-GAN is proposed to generate 2D facial images of arbitrary pose using the generated UV images.

6 (2017-08-17) Towards Large-Pose Face Frontalization in the Wild

Arxiv. Xilesou. Top/PDF / 1704.06…

Despite recent advances in facial recognition using deep learning, performance can be severely affected with large postural changes. Learning postures invariant features is one solution, but requires expensive large-scale data annotation and well-designed feature learning algorithms. In this paper, the 3D Morphable Model (3DMM) and GAN are combined to perform face transformation, which is called FF-GAN.

7 (2017-08-3) Beyond Face Rotation: Global and Local Perception GAN for Photorealistic and Identity Preserving Frontal View Synthesis

Arxiv. Xilesou. Top/PDF / 1704.04…

In this paper, a two-path generative adversal network (TP-GAN) is proposed to realize realistic frontal view synthesis by sensing both global structure and local details. In addition to the commonly used global encoder/decoder networks, four local block networks are proposed to deal with local textures. In addition, the loss combination of antagonism loss, symmetry loss and identity retention loss is introduced.

Welcome to pay attention to our official account: