A Deep Evaluator for Image Retargeting Quality by Geometrical and Contextual Interaction

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Image distortion during multi-device display impacts perceived quality.
• Existing objective quality assessment methods fail to accurately simulate human vision.
• Subjective evaluation is impractical for large-scale applications.

Purpose of the Study:

• To develop an accurate objective quality evaluation method for image retargeting.
• To simulate the human vision system's (HVS) perception of retargeted images.
• To improve upon current image retargeting quality assessment algorithms.

Main Methods:

• Proposed a deep quality evaluator using segmented stacked AutoEncoders (SAE).
• Employed regularization to mitigate overfitting in the deep learning framework.
• Developed two separate SAE models for geometrical shape and content matching.
• Combined scores from both models using weighting schemes.

Main Results:

• The proposed method demonstrates superior performance in evaluating image retargeting results.
• Achieved better performance compared to traditional methods across three benchmark databases.
• Effectively simulates human perception of retargeted images.

Conclusions:

• The deep quality evaluator based on SAE offers a more accurate assessment of image retargeting quality.
• This approach provides a viable alternative to subjective evaluations for large-scale systems.
• The method shows significant potential for improving image retargeting applications.