Block-Based CNN Image Processing for Compression Artifact Removal
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing convolutional neural networks (CNNs) struggle to effectively remove compression artifacts, such as blocking and ringing artifacts, at block boundaries due to the use of the same kernel for all pixels in an input image, and are not suitable for parallel processing.
Innovation Solution
An image processing apparatus and method that utilizes a convolutional neural network to extract feature information by dividing images into blocks, sampling pixels with the same location in each block, and applying different kernels to these blocks to remove compression artifacts, allowing for parallel processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If the same kernel is used for all pixels in an input image, then the CNN structure is simple and easy to implement, but compression artifacts at block boundaries cannot be effectively removed
Solution Approach 1:
The patent divides the input image into multiple blocks and extracts feature information at different locations within each block. By segmenting the image into blocks and processing different location groups separately, the system can apply appropriate kernels to specific regions, thereby effectively removing compression artifacts at block boundaries while maintaining overall processing simplicity.
Solution Approach 2:
The patent applies different kernels to different location groups within blocks based on their specific characteristics. Instead of using a single uniform kernel for the entire image, the system selects and applies appropriate kernels locally to different regions, enabling effective compression artifact removal in specific areas while maintaining computational efficiency.
2Manufacturing precision
If different filters are applied according to rules at block boundaries, then compression artifacts can be removed more effectively, but the implementation becomes complex and is not suitable for parallel processing
Solution Approach 1:
The patent segments the image into blocks and extracts feature information at different locations within each block. This segmentation allows the system to process different regions independently, which can be performed in parallel, thereby reducing implementation complexity while maintaining effective compression artifact removal.
Solution Approach 2:
The patent changes the kernel selection parameter based on the location of pixels within blocks. By using location information to determine which kernel to apply, the system can effectively remove compression artifacts without requiring complex rule-based filter selection, thus simplifying implementation while enabling parallel processing.
3Manufacturing precision
If different filters are applied according to rules at block boundaries, then compression artifacts can be removed more effectively, but parallel processing capability is lost
Solution Approach 1:
The patent divides the image into blocks and extracts feature information at different locations within each block. This segmentation creates independent processing units that can be handled simultaneously through parallel processing, thereby maintaining both compression artifact removal effectiveness and parallel processing capability.
Solution Approach 2:
The patent uses location-based parameter changes to determine kernel selection. Since this approach depends only on pixel location information rather than complex sequential rules, it can be efficiently implemented through parallel processing while still achieving effective compression artifact removal.
Data Source
AI summary
An image processing apparatus includes a memory storing at least one instruction; and a processor configured to execute the at least one instruction to use at least one neural network to: extract n pieces of first feature information from a first image, based on locations of pixels included in the first image, wherein n is a positive integer, generate n pieces of second feature information by performing a convolution operation between each of the n pieces of the first feature information and each of n kernels, and generate, based on the n pieces of the second feature information, a second image from which compression artifacts included in the first image are removed.