CNN Image Upsampling with Adaptive Resolution Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image signal processing methods are inflexible and computationally burdensome, unable to adjust up-sampling factors and process images in a variable pixel arrangement mode.
Innovation Solution
An image processing system incorporating a convolutional neural network (CNN) module, a combiner, and an activation module in a cascaded manner, which performs convolution operations to generate feature images and combines them into higher resolution images, allowing adaptive selection and amplification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard up-sampling modes (bicubic, linear) are used to increase image resolution, then image resolution is improved, but processing flexibility is lost
Solution Approach 1:
The patent implements dynamic up-sampling by using a convolutional neural network with multiple processing paths that can be adaptively selected. The system dynamically adjusts the up-sampling factor and processing mode based on input image characteristics, enabling both high resolution output and processing flexibility simultaneously. This resolves the contradiction by making the system adaptive rather than static.
Solution Approach 2:
The patent changes the parameter of up-sampling factor from fixed to variable. By implementing multiple processing paths with different up-sampling factors (e.g., 2x, 4x, 8x) and allowing dynamic selection among them, the system achieves both high resolution enhancement and flexibility in processing different image types and requirements.
2Measurement precision
If standard up-sampling modes are used to increase image resolution, then image resolution is improved, but computational burden increases
Solution Approach 1:
The patent segments the up-sampling process into multiple stages with different processing paths. Instead of applying a single computationally intensive algorithm to all images, the system divides processing into multiple manageable paths with varying complexity, allowing selective application based on needs. This reduces overall computational burden while maintaining high resolution output capability.
Solution Approach 2:
The patent implements partial processing by providing multiple processing paths where not all paths are executed for every image. The system selectively applies processing based on input characteristics and requirements, performing only the necessary computational steps rather than exhaustive processing, thereby reducing computational burden while achieving desired resolution enhancement.
3Device complexity
If fixed up-sampling factor is used, then processing simplicity is maintained, but adaptability to different resolution needs is lost
Solution Approach 1:
The patent creates a universal up-sampling system that handles multiple resolution requirements through a single integrated architecture. The convolutional neural network with multiple processing paths can perform various up-sampling tasks (2x, 4x, 8x, etc.) using the same system, making it multi-functional and adaptable to different resolution needs without requiring separate dedicated systems for each up-sampling factor.
Data Source
AI summary
An image processing system, an image processing method and a display device are provided. The image processing system includes at least one resolution conversion sub-system. The resolution conversion sub-system includes a CNN module, a combiner and an activation module connected in a cascaded manner. The CNN module is configured to perform convolution operation on an input signal to acquire a plurality of first feature images having a first resolution. The combiner is configured to combine the first feature images into a second feature image having a second resolution greater than the first resolution. The activation module is connected to the combiner and configured to perform a selection operation on the second feature image using an activation function.