Engineering Deep Representations for Modeling Aesthetic Perception.

Area of Science:

• Computer Vision
• Multimedia Analysis
• Machine Learning

Background:

• Existing aesthetic models lack region-level interpretability and adaptive feature engineering.
• Hand-crafted aesthetic criteria offer limited descriptiveness.

Purpose of the Study:

• Develop a deep architecture for learning aesthetically relevant visual attributes.
• Address shortcomings in current image aesthetic analysis models.

Main Methods:

• Utilized Flickr image tags for textual attribute discovery via a sparsity-constrained subspace algorithm.
• Employed weakly supervised learning to project textual attributes to image patches for visual attribute learning.
• Applied a five-layer convolutional neural network to mimic hierarchical human perception of visual concepts.

Main Results:

• Learned deep features demonstrate superiority in image retargeting, aesthetics ranking, and retrieval.
• Experimental results validate the effectiveness of the proposed deep architecture.

Conclusions:

• The developed deep architecture effectively learns localized visual attributes for aesthetic analysis.
• The approach offers improved descriptiveness and interpretability compared to traditional methods.