Improving the Harmony of the Composite Image by Spatial-Separated Attention Module.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Image composition is crucial in image processing, but inconsistencies between spliced regions and backgrounds reduce image quality.
• Existing methods often rely on direct neural network learning, which can be suboptimal for harmonizing composite images.

Purpose of the Study:

• To address the challenge of image harmonization by improving the stylistic consistency between spliced regions and their backgrounds.
• To develop a novel attention mechanism and framework for more effective image harmonization, including scenarios without explicit masks.

Main Methods:

• Propose a Spatial-Separated Attention Module (S2AM) to learn feature maps of masked and non-masked regions individually.
• Integrate the S2AM into a Unet-based framework at coarser low-level features through two distinct approaches.
• Extend the harmonization capability to composite images without requiring a specific mask.

Main Results:

• The proposed S2AM demonstrates superior performance compared to other state-of-the-art attention modules in image harmonization tasks.
• The developed framework shows significant advantages over existing image harmonization methods based on multiple evaluation criteria.

Conclusions:

• The S2AM is an effective module for improving image harmonization by addressing regional feature learning.
• The novel framework offers a significant advancement in harmonizing composite images, with potential applications beyond traditional masked harmonization.