Cascaded Decoding Units for High-Resolution Image Restoration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning techniques based on convolutional neural networks face challenges in processing high-resolution images due to limitations of small convolution kernels, requiring image down-sampling, which leads to information loss during compression and makes it difficult to restore the original high-resolution image effectively.
Innovation Solution
An image processing method and apparatus that utilizes cascaded decoding units with convolutional neural networks to process input images, performing inverse wavelet transforms and superimposition operations across multiple stages to enhance image resolution and recover lost information, while also incorporating training methods to optimize convolutional networks for improved image encoding and decoding.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If image down-sampling is performed to enable processing by convolutional neural networks, then images can be processed by the network, but information loss occurs during compression
Solution Approach 1:
The patent segments the image processing into multiple stages with different convolution kernel sizes. Small kernels (3×3) are used for initial feature extraction, while large kernels (17×17, 25×25, or larger) are used in subsequent stages to capture broader contextual information and restore lost details, thereby reducing information loss while maintaining network processability
Solution Approach 2:
The patent transitions from traditional single-scale convolution to multi-scale convolution by introducing kernels of different sizes across multiple stages. This dimensional change in kernel size allows the network to process images at different receptive fields, enabling both efficient processing and recovery of compressed information
2Manufacturing precision
If high-resolution images are processed directly, then image quality is maintained, but convolutional neural networks with small kernels cannot perceive the entire image
Solution Approach 1:
The patent employs dynamic kernel size adjustment across processing stages. The network adapts from small kernels in early stages to progressively larger kernels in later stages, allowing the system to dynamically match the receptive field to the processing requirements at each stage, thereby maintaining both resolution quality and network perception capability
Solution Approach 2:
The patent implements a nested multi-stage architecture where each processing stage builds upon the previous one. Small kernel processing results are fed into larger kernel processing, creating a nested structure where larger kernels encompass the processing of smaller kernels, enabling progressive perception of the entire high-resolution image
3Productivity
If multiple small images are generated from high-resolution images, then convolutional neural networks can process them, but the converted images have lower resolution and smaller file sizes
Solution Approach 1:
The patent merges multiple processing streams with different kernel sizes and combines their outputs through concatenation and fusion operations. This combining approach allows the network to leverage both the efficiency of small-kernel processing and the quality-preserving capabilities of large-kernel processing, achieving high-resolution output with improved processing efficiency
Data Source
AI summary
An image processing apparatus and method, an image processing system and a training method are disclosed. The image processing method comprises: receiving an input image; and inputting the input image to K stages of cascaded decoding units, to process the input image to obtain an output image, wherein an ith stage of decoding unit receives mK+1−i input images and outputs mK−i output images, a resolution of the output images is greater than a resolution of the input images, and the number of image components of the output images is less than the number of image components of the input images, where K, i and m are positive integers and 1≤i≤K.