Theme-Aware Aesthetic Distribution Prediction With Full-Resolution Photographs.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Aesthetic quality assessment (AQA) is challenging due to complex aesthetic factors.
• Current deep neural network (DNN) methods for AQA require fixed-size inputs, often damaging features through transformations like resizing or cropping.
• Existing adaptive pooling methods also struggle to fully capture aesthetic features from fixed-size inputs.

Purpose of the Study:

• To propose a novel method for full-resolution image AQA that overcomes the limitations of fixed-size input transformations.
• To address the issue of theme criterion bias, where aesthetic evaluations vary across different themes.
• To improve the accuracy and robustness of AQA models.

Main Methods:

• A full-resolution AQA method combining image padding with region of image (RoM) pooling.
• Encoding and fusing image aspect ratios with visual features to mitigate shape information loss from RoM pooling.
• Developing a theme-aware model that incorporates theme information to guide predictions, using an attention-based feature fusion module.

Main Results:

• The proposed RoM pooling effectively pools image features while discarding padded regions, minimizing side effects.
• Encoding aspect ratios and fusing them with visual features helps recover shape information.
• The theme-aware model and attention-based fusion module demonstrated significant improvements in AQA accuracy.
• Extensive experiments confirmed the superiority of the proposed method over state-of-the-art techniques.

Conclusions:

• The proposed method achieves effective full-resolution AQA by preserving image features and addressing shape information loss.
• Incorporating theme information and aspect ratios significantly enhances AQA performance.
• The developed approach offers a robust and accurate solution for aesthetic quality assessment.