Convolutional Neural Network for Full-Reference Image Quality Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image quality assessment methods for digital images, especially Full-Reference IQA, struggle to accurately align with human perception across different types of distortions and are not efficiently objective, particularly when reference images are available.
Innovation Solution
A convolutional neural network system that processes both distorted and reference images to produce a quality metric, trained using human perception-based similarity scores and adjusted parameters to improve correlation with human visual perception.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional FR-IQA metrics (MSE, PSNR) are used to measure image quality, then computational efficiency is improved, but correlation with human perception deteriorates
Solution Approach 1:
The patent replaces conventional pixel-by-pixel comparison methods (MSE, PSNR) with a convolutional neural network that learns perceptual features directly from image data. The CNN substitutes traditional mathematical metrics with a learned model that captures human visual perception characteristics, achieving both efficiency and perceptual accuracy.
Solution Approach 2:
The patent transforms the approach from using fixed mathematical parameters (MSE, PSNR) to using learned features from a trained neural network. The CNN extracts relevant visual features and transforms them into quality metrics that better align with human perception while maintaining computational efficiency through automated feature extraction.
2Measurement precision
If bottom-up approaches model the human visual system to improve perception alignment, then correlation with human perception is improved, but system complexity increases
Solution Approach 1:
The patent employs a convolutional neural network that automatically learns and adapts to model human visual perception characteristics through training data. Instead of manually encoding complex HVS properties, the system self-learns the relevant features and relationships, reducing the need for explicit complex modeling while achieving perceptual accuracy.
3Productivity
If low-level visual features are used for quality assessment, then computational efficiency is improved, but performance varies across different distortion types
Solution Approach 1:
The patent implements a dynamic feature extraction system using a trained convolutional neural network that adapts its feature representation based on the input image and distortion type. The CNN learns to dynamically adjust which features are most relevant for different distortion categories, maintaining both efficiency and consistent performance across diverse distortion scenarios.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Embodiments generally relate to providing systems and methods for assessing image quality of a distorted image relative to a reference image. In one embodiment, the system comprises a convolutional neural network that accepts as an input the distorted image and the reference image, and provides as an output a metric of image quality. In another embodiment, the method comprises inputting the distorted image and the reference image to a convolutional neural network configured to process the distorted image and the reference image and provide as an output a metric of image quality.