Gaming Super Resolution Network Architecture for Image Detail Preservation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional super-resolution techniques produce blurry and corrupted images due to the lack of utilization of non-linear information and fail to preserve original image details during upscaling, as they are generalizable and lack specific knowledge of immediate problems.
Innovation Solution
The proposed solution involves a Gaming Super Resolution (GSR) network architecture that employs a combination of linear and non-linear up-sampling operations, using convolutional layers and activation functions to efficiently super-resolve images, preserving original information and improving fidelity by training weights with a corpus of images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional super-resolution techniques are used, then image upscaling is achieved, but image quality becomes blurry and corrupted due to lack of non-linear information utilization
Solution Approach 1:
The patent changes the mathematical parameters of the upscaling operation by switching from linear to non-linear transformations. The non-linear activation functions (ReLU, Leaky ReLU, ELU, GELU, Swish) enable the network to capture complex relationships and preserve original image details that linear methods cannot capture, thereby resolving the contradiction between upscaling and maintaining image quality.
Solution Approach 2:
The patent employs a composite approach by combining multiple non-linear activation functions within the neural network architecture. This composite combination allows the system to leverage the strengths of different activation functions to better preserve original image information while achieving high-quality upscaling, addressing both the upscaling requirement and the detail preservation requirement simultaneously.
2Manufacturing precision
If deep learning approaches are used, then super-resolution is achieved, but color and detail information is lost due to lack of original image incorporation
Solution Approach 1:
The patent implements feedback mechanisms where the network continuously references and incorporates original image information during the upscaling process. The non-linear activation functions enable the network to feedback original color and detail information throughout the computational graph, ensuring that this information is preserved rather than lost, thus resolving the contradiction between achieving super-resolution and maintaining original information.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the original image to enhance color and detail information before feeding it into the neural network. This preliminary enhancement ensures that the original image information is amplified and preserved throughout the upscaling process, preventing loss of color and detail while achieving high-quality super-resolution.
3Adaptability or versatility
If conventional neural network architectures are used, then generalizable processing is achieved, but specific knowledge of immediate problems is lacking
Solution Approach 1:
The patent applies local quality by enabling the neural network to adapt its non-linear transformation parameters locally for different types of image content and problems. The non-linear activation functions allow the network to adjust its behavior based on the specific characteristics of the input image, providing problem-specific accuracy while maintaining generalizability through the universal application of non-linear principles across different tasks.
Data Source
AI summary
A processing device is provided which includes memory and a processor. The processor is configured to receive an input image having a first resolution, generate at least one linear down-sampled version of the input image via a linear upscaling network, generate at least one non-linear down-sampled version of the input image via a non-linear upscaling network, extract a first feature map from the at least one linear down-sampled version of the input image, and extract a second feature map from the at least one non-linear down-sampled version of the input image. The processor is also configured to convert the at least one linear down-sampled version of the input image and the at least one non-linear down-sampled version of the input image into pixels of an output image having a second resolution higher than the first resolution using the first feature map and the second feature map.