Multispectral Image Restoration with GAN
summary
Presents a restoration pipeline that uses a trained GAN (implemented in PyTorch) to inpaint random-size, random-location corruptions in 4-channel multispectral aerial tiles by optimizing the latent z for each corrupted image instead of retraining the generator. Evaluated on a sugar‑beet multispectral dataset from Switzerland and Germany, the method recovers texture and produces visually acceptable but slightly blurry results with noticeable seams.
AI-written summary · about
chapters (12)
- 0:00Start
- 0:38Black and white images
- 0:58Color images
- 1:38How do we work with them?
- 3:09Applications
- 7:19Part 1. Training the discriminator
- 8:21Part 2. Training the generator
- 9:09Basic trained GAN
- 9:49Dataset used
- 10:59Stage 1: Training the basic GAN structure
- 11:23Stage 2. Restoration with trained GAN
- 12:13Code
related talks
12mENHANCE! Upscaling Images with Neural Networks
18mDance of the Ancestors...
36mSemi-supervised deep learning with GANs
33mLow to High Resolution
52mDeep Generative Models And Unsupervised Methods For Inverse Problems
32mEnsembles of GANs as a Data Augmentation Technique for Alzheimer research
37mHarnessing the power of Generative Adversarial Networks (GANS) for supervised learning
25mSurfzone Topography-informed Deep Learning Techniques to Nearshore Bathymetry with Sparse Data