Exploring Rich Subjective Quality Information for Image Quality Assessment in the Wild

Overview

• Traditional in-the-wild image quality assessment (IQA) models are trained on mean opinion scores (MOS) but miss rich subjective quality information like standard deviation of opinion scores (SOS) and distribution of opinion scores (DOS).
• This paper proposes a novel IQA method called RichIQA that explores this rich subjective rating information to predict image quality.
• RichIQA has two key novel designs: a three-stage image quality prediction network and a multi-label training strategy using MOS, SOS, and DOS.

Plain English Explanation

RichIQA is a new way to assess the quality of images found in the real world. Traditional methods only look at the average opinion score, but RichIQA also considers the spread and distribution of opinions.

The three-stage network in RichIQA is designed to mimic how the human brain processes images. It has powerful feature extraction capabilities and can remember both short-term and long-term information.

RichIQA also uses a special multi-label training strategy. Instead of just training on the average opinion score, it also uses the spread of opinions and the full distribution of scores. This helps the network make better predictions and be more reliable.

By considering this rich subjective information, RichIQA can predict the full distribution of image quality, not just a single score. This provides more detailed and useful feedback compared to traditional methods.

Technical Explanation

The key innovations in RichIQA are:

1. Three-stage Quality Prediction Network: This network uses the powerful feature representation of Convolutional Vision Transformers (CvT) and mimics the short-term and long-term memory mechanisms of the human brain to predict image quality.

2. Multi-label Training Strategy: The network is trained concurrently on MOS, SOS, and DOS subjective quality information, allowing it to fully exploit the rich rating data.

The three-stage architecture is designed to capture different aspects of image quality. The first stage extracts low-level visual features, the second stage models short-term dependencies, and the third stage models long-term dependencies. This hierarchical structure allows the network to make sophisticated quality predictions.

The multi-label training strategy ensures the network learns from the full range of subjective quality information, not just the average opinion score. This enhances the model's performance and generalization ability compared to approaches that only use MOS.

Critical Analysis

The paper provides a thorough evaluation of RichIQA on multiple large-scale in-the-wild IQA datasets, demonstrating its strong performance compared to state-of-the-art methods.

However, the authors do not discuss any limitations or potential issues with their approach. For example, the complexity of the three-stage network may make it computationally expensive or difficult to deploy in real-world applications.

Additionally, the authors could have explored how RichIQA's ability to predict quality distributions, rather than just single scores, could benefit end users. Further research is needed to understand the practical implications of this rich quality information.

Conclusion

RichIQA is a novel IQA method that leverages the full range of subjective quality ratings, including opinion score distribution, to make more nuanced and reliable predictions of image quality. Its three-stage network architecture and multi-label training strategy are key innovations that allow it to outperform existing approaches.

This research highlights the importance of considering rich subjective information beyond just average opinion scores when assessing image quality, particularly for real-world applications. RichIQA's ability to capture the full distribution of quality perceptions could lead to significant improvements in areas like image optimization, quality control, and user experience design.