Ranking-Preserving Cross-Source Learning for Image Retargeting Quality Assessment.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image retargeting aims to resize images while preserving visual quality, attracting significant research interest.
• Objective Quality Assessment (OQA) methods are crucial for automatically evaluating retargeting results.
• Existing OQA methods struggle with consistency when comparing results from different source images.

Purpose of the Study:

• To develop a learning-based OQA method that overcomes the limitations of existing approaches.
• To improve the accuracy and generalizability of objective quality assessment for image retargeting.
• To provide a reliable method for ranking retargeting results, even across different source images.

Main Methods:

• A General Regression Neural Network (GRNN) model was trained using relative quality scores from same-source image retargeting results.
• A novel training scheme with provable convergence was developed to learn a common base scalar for same-source results.
• The model was trained and evaluated on human preference data from the RetargetMe benchmark.

Main Results:

• The proposed GRNN model demonstrated superior ranking prediction compared to ten representative OQA methods.
• The method achieved better generalizability across different datasets, outperforming existing approaches.
• The learned source-specific offsets enabled effective comparison of retargeting results from different source images.

Conclusions:

• The proposed learning-based OQA method effectively addresses the challenges of assessing image retargeting quality.
• The approach offers improved accuracy, ranking preservation, and generalizability for objective quality assessment.
• This work provides a more robust and reliable solution for automated evaluation of image retargeting techniques.