20230087476. METHODS AND APPARATUSES FOR PHOTOREALISTIC RENDERING OF IMAGES USING MACHINE LEARNING simplified abstract (KWAI INC.)
METHODS AND APPARATUSES FOR PHOTOREALISTIC RENDERING OF IMAGES USING MACHINE LEARNING
Organization Name
Inventor(s)
Oliver Dayun Liu of Palo Alto CA (US)
Haibin Huang of Palo Alto CA (US)
Chongyang Ma of Palo Alto CA (US)
METHODS AND APPARATUSES FOR PHOTOREALISTIC RENDERING OF IMAGES USING MACHINE LEARNING - A simplified explanation of the abstract
This abstract first appeared for US patent application 20230087476 titled 'METHODS AND APPARATUSES FOR PHOTOREALISTIC RENDERING OF IMAGES USING MACHINE LEARNING
Simplified Explanation
The abstract describes a method for training a neural network and processing images using this method. The method involves obtaining unpaired images from different domains, scaling the images, cropping them into training patches, inputting the patches into the neural network, and calculating a contrastive loss based on selected sub-patches. The model parameters of the neural network are then updated based on the contrastive loss and a generative adversarial network loss.
- The method involves training a neural network using unpaired images from different domains.
- The images are scaled and cropped into training patches with different contents.
- The training patches are inputted into the neural network to generate output patches.
- A contrastive loss is calculated based on selected sub-patches from the training patches and a corresponding positive sub-patch from the output patches.
- The model parameters of the neural network are updated based on the contrastive loss and a generative adversarial network loss.
Potential Applications
- Image processing and enhancement
- Style transfer and image translation
- Data augmentation for training neural networks
Problems Solved
- Lack of paired images for training neural networks
- Difficulty in training neural networks for image processing tasks
- Limited availability of diverse training data
Benefits
- Allows training of neural networks using unpaired images
- Enables image processing and enhancement without paired images
- Increases the diversity and quality of training data for neural networks
Original Abstract Submitted
a neural network training method, an image processing method, and apparatuses thereof are provided. the neural network training method includes obtaining a first domain image and a second domain image, where the first domain image and the second domain image are unpaired images in different domains; obtaining a scaled first domain image by scaling, at an iteration, the first domain image; obtaining a training patch by cropping the scaled first domain image, where each training patch has a same number of pixels with different contents; inputting the training patch into the neural network at the iteration, and outputting an output patch; calculating a contrastive loss based on a query sub-patch and negative sub-patches selected from the training patch and a corresponding positive sub-patch selected from the output patch; and updating model parameters of the neural network based on the contrastive loss and a generative adversarial network loss.
