Image Quality Assessment Using Contrastive Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Traditional image quality assessment (IQA) often requires large labeled datasets.
• Developing robust IQA models that generalize across diverse distortions remains a challenge.

Purpose of the Study:

• To develop a self-supervised framework for learning image quality representations.
• To create a deep learning model for no-reference (NR) image quality evaluation without subjective labels.

Main Methods:

• Utilized distortion type and degree prediction as an auxiliary task for self-supervised feature learning.
• Trained a Convolutional Neural Network (CNN) using a contrastive pairwise objective.
• Developed the CONTRastive Image QUality Evaluator (CONTRIQUE) model.

Main Results:

• CONTRIQUE achieved competitive performance against state-of-the-art NR IQA models.
• Learned representations demonstrated robustness and generalization across synthetic and realistic distortions.
• Effective quality representations were obtained without extensive labeled subjective datasets.

Conclusions:

• Self-supervised learning via distortion prediction is a viable approach for IQA.
• CONTRIQUE offers a powerful and generalizable solution for NR image quality assessment.
• The findings reduce the reliance on large, manually annotated subjective image quality datasets.