Proxy Network Training for Image Generative Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image and video compression technologies face challenges in minimizing distortions and artifacts during lossy compression, particularly in training image generative networks, where existing methods struggle with stability and robustness against adversarial samples and artifact generation.
Innovation Solution
A computer-implemented method and system for training an image generative network using a proxy network that evaluates a gradient intractable perceptual metric across multiple scales, allowing for non-differentiable metric training and improved stability and robustness against adversarial samples and artifacts, by generating output images without tracking gradients and optimizing network parameters through backpropagation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If lossy compression is applied to images, then data transmission efficiency is improved, but image quality deteriorates with distortions and artifacts
Solution Approach 1:
A proxy network is introduced as an intermediary to approximate the gradient intractable perceptual metric. This proxy network provides differentiable gradients that can be used for training the image generative network, enabling optimization toward perceptual quality without directly computing the non-differentiable metric during backpropagation.
Solution Approach 2:
The patent employs multi-scale training by processing images at different resolutions and aggregating losses across scales. This parameter change approach allows the network to learn both fine details and global structures, improving perceptual quality while maintaining compression efficiency.
2Manufacturing precision
If gradient intractable perceptual metrics are used for training, then image quality evaluation is improved, but training stability deteriorates
Solution Approach 1:
The proxy network serves as a differentiable intermediary that approximates the gradient intractable perceptual metric. By training the proxy network separately and using it to provide gradients during main network training, the system achieves both accurate perceptual evaluation and training stability.
Solution Approach 2:
The proxy network is trained beforehand on a dataset to learn the mapping between image pairs and perceptual metric values. This preliminary action prepares the gradient approximation mechanism before actual network training, ensuring stable gradient flow during the main training process.
3Ease of operation
If standard training methods are used, then training simplicity is maintained, but robustness against adversarial samples deteriorates
Solution Approach 1:
The proxy network acts as a robust intermediary that provides stable gradient approximations resistant to adversarial perturbations. By decoupling the non-differentiable perceptual metric computation from the gradient computation, the system achieves robustness while maintaining training simplicity.
Data Source
AI summary
A computer-implemented method of training an image generative network fθ for a set of training images, in which an output image {circumflex over (x)} is generated from an input image x of the set of training images non-losslessly, and in which a proxy network is trained for a gradient intractable perceptual metric that evaluates a quality of an output image {circumflex over (x)} given an input image x, the method of training using a plurality of scales for input images from the set of training images. In an embodiment, a blindspot network bα is trained which generates an output image {tilde over (x)} from an input image x. Related computer systems, computer program products and computer-implemented methods of training are disclosed.