AI Image Filtering Using Motion-Based Reference Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional rule-based filtering algorithms in image codecs like H.264 AVC, HEVC, and VVC struggle to handle the increasing resolution and diversified content of images, necessitating an AI-based filtering approach that adapts to image characteristics.
Innovation Solution
An AI-based filtering method using a neural network that processes motion information to determine reference blocks, reducing errors between reconstructed and original images by training on input data including current and reference blocks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If rule-based filtering algorithms are used, then filtering performance is maintained for traditional images, but the algorithms cannot adapt to increasing image resolutions and diversified content
Solution Approach 1:
The patent replaces traditional rule-based filtering mechanisms with an AI-based filtering system that uses neural networks to automatically learn and adapt filtering parameters. The neural network is trained on diverse image data including different resolutions and content types, enabling it to generalize across various image characteristics without requiring explicit rules for each scenario.
Solution Approach 2:
The patent dynamically adjusts filtering parameters based on image characteristics by using a neural network that takes image data as input and outputs optimized filtering parameters. The system changes parameters such as filtering strength, kernel size, and other processing parameters adaptively according to the specific image resolution and content type, rather than using fixed rules.
2Adaptability or versatility
If AI-based filtering is implemented to adapt to diverse image content, then adaptability improves, but computational complexity and processing time increase
Solution Approach 1:
The patent performs preliminary training of the neural network offline using large datasets containing diverse image resolutions and content types. This pre-training phase captures general patterns and characteristics, allowing the model to make accurate predictions during actual filtering operations without requiring complex real-time computations. The pre-trained model serves as a ready-to-use filter that adapts quickly to new images.
Solution Approach 2:
The patent applies selective filtering based on image characteristics, using the full AI-based filtering capability only when needed for complex or high-resolution images. For simpler or lower-resolution images, the system may use simplified filtering or skip certain processing stages, thereby reducing computational overhead while maintaining adaptability where required.
3Measurement precision
If traditional rule-based algorithms are used, then processing speed is maintained, but error reduction between reconstructed and original images is limited
Solution Approach 1:
The patent replaces traditional rule-based error reduction mechanisms with a neural network that learns optimal filtering strategies from training data. The neural network analyzes the relationship between reconstructed images and original images, automatically identifying patterns that minimize errors across diverse image types and resolutions, thereby achieving superior error reduction compared to fixed rules.
Solution Approach 2:
The patent implements a feedback mechanism where the neural network is trained using ground truth data from original images to continuously improve its filtering performance. The system compares filtered output with original images, uses the difference as feedback to adjust parameters and learn better filtering strategies, thereby progressively reducing errors while maintaining efficient processing through the trained model.
Data Source
AI summary
An image processing method for artificial intelligence (AI)-based filtering includes: obtaining motion information regarding inter prediction of a current block of a current reconstructed image, based on the motion information regarding the inter prediction, determining at least one reference reconstructed image regarding the current reconstructed image, based on the motion information and the at least one reference reconstructed image, determining at least one reference block, by inputting the current block and the at least one reference block to a neural network, obtaining an output block corresponding to the current block, and based on the output block, obtaining a modified current reconstructed image, wherein the neural network is trained using training data including the current block, the at least one reference block, and the output block, to reduce an error between the current block and an original block corresponding to the current block.