vaknafusu / gan-imagerepairing

基于深度生成对抗网络gan的图像修复模型

DcGAN人脸修复

My individual project as a B.S student in Tongji University.

Deep learning for image inpainting

Environment

• Tensorflow 1.0.0
• Cuda 8.0
• Cudnn 5.1
• python 2.7
• Openface
• Scipy
• Pillow
• Numpy
• OpenCv
• Dlib

Flow chart

Dataset

• CelebA: 10177 people’s 20599 face images, for face images training
• LFW: 5000 people’s 10000+ face images, for face image testing
• Oxford Building: 5062 buildings, in which 1024 as testset, others are trainset
• INRIA: 1491 images of different kind, in which 496 as testset, others are trainset

Assessment Value

Use the following values to compare the similarity of original images and completed images, which can tell the performance of repairing.

• MSE(Mean Square Error)
• PSNR(Peak signal-to-noise ratio)
• SSIM(Structural Similarity)

Some test results on LFW

Some test results on nature images (INRIA and Oxford Buildings)

Different damaged area

• different damaged type

• different damaged size

• And so on...

Expections

Here are some expections during my master career at King's College London

• optimise model
• build an API and deploy online