Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
157 tokens/sec
GPT-4o
43 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
4 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Perceptual Quality Prediction on Authentically Distorted Images Using a Bag of Features Approach (1609.04757v1)

Published 15 Sep 2016 in cs.CV

Abstract: Current top-performing blind perceptual image quality prediction models are generally trained on legacy databases of human quality opinion scores on synthetically distorted images. Therefore they learn image features that effectively predict human visual quality judgments of inauthentic, and usually isolated (single) distortions. However, real-world images usually contain complex, composite mixtures of multiple distortions. We study the perceptually relevant natural scene statistics of such authentically distorted images, in different color spaces and transform domains. We propose a bag of feature-maps approach which avoids assumptions about the type of distortion(s) contained in an image, focusing instead on capturing consistencies, or departures therefrom, of the statistics of real world images. Using a large database of authentically distorted images, human opinions of them, and bags of features computed on them, we train a regressor to conduct image quality prediction. We demonstrate the competence of the features towards improving automatic perceptual quality prediction by testing a learned algorithm using them on a benchmark legacy database as well as on a newly introduced distortion-realistic resource called the LIVE In the Wild Image Quality Challenge Database. We extensively evaluate the perceptual quality prediction model and algorithm and show that it is able to achieve good quality prediction power that is better than other leading models.

Citations (284)

Summary

  • The paper presents FRIQUEE, a novel method that predicts image quality using a bag-of-features approach on authentically distorted images.
  • It employs divisive normalization and multi-domain transformations across color spaces to extract quality-aware features that mirror human visual perception.
  • The model significantly outperforms traditional IQA methods on the LIVE In the Wild dataset, demonstrating superior correlation with human opinion scores.

Perceptual Quality Prediction on Authentically Distorted Images

The paper by Ghadiyaram and Bovik presents a novel approach for predicting the perceptual quality of authentically distorted images, which often contain complex, composite distortions. Traditional image quality assessment (IQA) models tend to focus on synthetically distorted images that do not adequately reflect the comprehensive distortion types encountered in real-world scenarios. The authors propose a robust framework utilizing a bag of features approach that leverages natural scene statistics (NSS) across various perceptual color spaces and transform domains to evaluate image quality.

The main contribution of the paper is the introduction of a perceptually motivated bag of features method designed to circumvent the limitations of existing top-performing blind IQA models which primarily handle inauthentic distortions. This approach extracts a broad array of quality-aware features from real-world authentically distorted images, enabling the training of a regressor to predict image quality in a way that closely aligns with human visual judgment. Utilizing a large dataset containing authentically distorted images, the paper emphasizes the pivotal differences between synthetic and authentic distortions, highlighting the destabilization of statistical regularities assumed by prior models.

Among the methodological innovations, the paper proposes several new statistical feature maps that enhance the discriminative power of the model. These include divisive normalization techniques across multiple color spaces (RGB, CIELAB, and LMS) and transformations such as the steerable pyramid wavelet transform. These transformations are theorized to model processes occurring in the human visual system, particularly in the retina and cortex.

Numerically, the proposed model, named FRIQUEE (Feature Maps based Referenceless Image QUality Evaluation Engine), demonstrates significant improvements in correlation with human opinion scores over existing methods when tested against the LIVE In the Wild Image Quality Challenge Database. The model achieves superior prediction accuracy compared to other blind IQA models like BRISQUE, DIIVINE, and BLIINDS-II, illustrating its robustness in handling real-world distortions.

Furthermore, the paper carefully outlines the implications of utilizing authentic distortion datasets, underscoring the necessity for IQA models to integrate features that capture more diverse distortion characteristics. It postulates that further investigation into authentic distortions will be instrumental in developing more robust quality predictors. The research also anticipates implications for practical applications, such as enhancing consumer devices' image capture and processing capabilities.

In terms of future work, the authors suggest extending their methodology to video quality assessment, leveraging the framework developed for static images to accommodate temporal dynamics present in video data. Additionally, the paper invites exploration into practical implementations that optimize image quality directly within digital cameras and mobile devices.

Overall, this work represents a noteworthy progression in the domain of perceptual image quality assessment, providing a comprehensive framework that significantly enhances the capability of predicting the quality of authentically distorted images—those most representative of common user experiences.