A Deep Network Solution for Attention and Aesthetics Aware Photo Cropping

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Automatic photo cropping is crucial for preserving image content and aesthetic appeal.
• Existing methods often struggle with balancing information preservation and visual quality.

Purpose of the Study:

• To develop an efficient deep learning model for high-quality photo cropping.
• To address the challenge of limited training data in image cropping tasks.

Main Methods:

• A novel two-branch neural network combining attention box prediction (ABP) and aesthetics assessment (AA).
• Shared convolutional feature maps for computational efficiency between the ABP and AA sub-networks.
• Generation of cropping candidates followed by aesthetic quality selection.

Main Results:

• The model effectively preserves important image parts while ensuring aesthetic pleasantness.
• High-quality cropping results were achieved, validated on benchmark datasets.
• The approach demonstrates computational efficiency, running at 5 frames per second.

Conclusions:

• The proposed attention-based deep learning model significantly improves automatic photo cropping.
• The dual-branch architecture offers an effective and efficient solution for image cropping.
• The method shows promise for applications requiring rapid, high-quality image cropping.