No-Reference Image Quality Assessment for Low-Latency Capture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image quality assessment methods are subjective and costly, and traditional signal fidelity measures do not align well with human visual perception, making it challenging to objectively evaluate image quality without a reference image.
Innovation Solution
A no-reference (NR) framework using a deep neural network that predicts image quality scores by feeding pixels into a convolutional neural network, trained on human-scored datasets, to estimate image quality factors like exposure, blurriness, and color tone, and runs asynchronously to avoid latency in the image capture pipeline.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional signal fidelity measures are used to assess image quality, then the assessment process is simple and fast, but the assessment results do not align well with human visual perception
Solution Approach 1:
The patent introduces a no-reference image quality assessment model as an intermediary system that bridges the gap between simple signal fidelity measures and human visual perception. This model processes image data and generates quality scores that correlate with human evaluation without requiring complex reference images or subjective human assessment, thereby improving measurement precision while controlling system complexity.
Solution Approach 2:
The patent replaces traditional mechanical/manual image quality assessment methods with an automated deep learning-based no-reference model. This substitution eliminates the need for complex reference image comparisons and subjective human evaluation, providing objective quality scores that align with human perception while simplifying the overall assessment system.
2Productivity
If no-reference image quality assessment is implemented, then scalability and objectivity are improved, but computational complexity increases
Solution Approach 1:
The patent extracts the essential quality assessment functionality into a standalone no-reference model that operates independently without requiring reference images or complex external systems. This extraction enables the model to be deployed at scale across multiple devices and platforms, improving productivity and scalability while managing computational complexity through focused feature processing.
3Measurement precision
If deep neural networks are used for image quality assessment, then assessment accuracy improves, but inference cost and latency increase
Solution Approach 1:
The patent employs a deep neural network model that has been pre-trained on extensive image datasets before deployment. This preliminary training action allows the model to achieve high assessment accuracy while reducing inference time during actual use, as the complex learning processes have already been completed during the pre-training phase.
Solution Approach 2:
The patent optimizes the deep neural network model by adjusting parameters such as model architecture, input image dimensions, and processing batch sizes to balance accuracy and inference speed. These parameter changes enable the system to maintain high measurement precision while minimizing inference latency and computational cost.
Data Source
AI summary
A method of image quality assessment, performed by one or more processors in an image capture device, is disclosed. The method comprising receiving notification of capture of an image and in response, initiating an image assessment task to assess the quality of the image. The assessment task comprises determining suitability of the image for image quality assessment, running an image quality assessment model on the image to generate image quality assessment results, collecting data related to the capture of the image, and transmitting the results to an image quality assessment repository. The image assessment task may be a lower priority asynchronous task.