KonIQ-10k: An ecologically valid database for deep learning of blind image quality assessment

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Existing image quality assessment (IQA) models are hindered by limited dataset sizes, requiring significant resources for data generation and annotation.
• Developing large-scale, ecologically valid IQA datasets is crucial for advancing deep learning models.

Purpose of the Study:

• To present a systematic and scalable method for creating KonIQ-10k, the largest IQA dataset to date.
• To introduce a novel deep learning model, KonCept512, for improved IQA performance and generalization.
• To establish a new benchmark for in-the-wild IQA with authentic distortions and diverse content.

Main Methods:

• Creation of KonIQ-10k dataset: 10,073 images with quality scores obtained via crowdsourcing (1.2 million ratings from 1,459 workers).
• Development of KonCept512: A deep learning model based on InceptionResNet architecture, trained at a higher resolution (512 × 384).
• Evaluation of model generalization on the LIVE-in-the-Wild dataset.

Main Results:

• KonIQ-10k established as the largest IQA dataset with high ecological validity.
• KonCept512 achieved a high generalization performance (0.921 SROCC) on its test set.
• KonCept512 demonstrated strong performance on the LIVE-in-the-Wild dataset (0.825 SROCC), comparable to 9 subjective scores.

Conclusions:

• The developed KonIQ-10k dataset and KonCept512 model significantly advance the field of image quality assessment.
• The findings pave the way for more generalizable and accurate deep learning-based IQA models.
• Crowdsourcing provides a viable method for creating large-scale, high-quality datasets for machine learning research.