Folded Generator Training for Faster Image Resolution Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current generative adversarial network (GAN) techniques for improving image resolution face challenges such as slow training and unsatisfactory reasoning effects, particularly in the context of mobile terminals and high-resolution image processing.
Innovation Solution
A method involving a folding operation on the weights of convolutional layers of a pre-trained generator to create a folded convolutional layer, which is then embedded into an image processing model and trained using pairs of sample images with differing resolutions, enabling efficient training and improved resolution enhancement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional GAN techniques are used for image resolution enhancement, then image generation capability is achieved, but training time and computational resources are excessive
Solution Approach 1:
The patent applies preliminary action by pre-training a generator model separately before embedding it into the final image processing model. This pre-training phase allows the generator to learn effective image generation capabilities in advance, so that during the subsequent training phase with paired images, the model converges faster and requires fewer computational resources to achieve the desired resolution enhancement performance.
2Measurement precision
If conventional GAN techniques are used for image resolution enhancement, then image generation capability is achieved, but computational resources and complexity are excessive
Solution Approach 1:
The patent applies segmentation by dividing the training process into distinct phases: first training the generator model independently, then embedding it into the image processing model for paired-image training. This segmentation of the training process reduces the overall complexity by breaking down the computationally intensive task into manageable stages, each with specific objectives and resource requirements.
Solution Approach 2:
By performing preliminary training of the generator before integration into the full model, the patent reduces the complexity of the overall training process. The pre-trained generator serves as a ready-made component that requires fewer adjustments during final training, thereby reducing computational resources and simplifying the training workflow.
3Adaptability or versatility
If GAN is deployed in edge devices, then image processing capability is provided, but security risks and latency increase
Solution Approach 1:
The patent employs copying by training the generator model separately and then embedding its weights into the image processing model. This creates a streamlined architecture where the pre-trained generator can be efficiently deployed on edge devices. The copied and integrated model structure reduces computational overhead and latency while maintaining security by avoiding the need to deploy the entire complex GAN architecture on resource-constrained edge devices.
Data Source
AI summary
Embodiments of the present disclosure relate to a method, an electronic device, and a computer program product for training an image processing model. The method in an illustrative embodiment includes: obtaining a folding weight of a folded convolutional layer of a pre-trained generator by performing a folding operation on a plurality of weights of a plurality of convolutional layers of the pre-trained generator. The method further includes: embedding the pre-trained generator into the image processing model. The method further includes: training the image processing model using a plurality of pairs of sample images, wherein at least one pair of sample images of the plurality of pairs of sample images includes a first sample image having a first resolution and a second sample image having a second resolution, and wherein the first resolution is less than the second resolution.