Frequency-Aware Feature Fusion for Dense Image Prediction.

Area of Science:

• Computer Vision
• Deep Learning
• Image Processing

Background:

• Dense image prediction requires high-resolution features with strong category information and precise spatial details.
• Current hierarchical models often fuse features by adding upsampled deep features with low-level high-resolution features.
• This direct fusion can lead to intra-category inconsistency and blurred object boundaries due to disturbed high-frequency information.

Purpose of the Study:

• To address the limitations of traditional feature fusion methods in dense image prediction tasks.
• To propose a novel feature fusion technique that improves feature consistency and boundary definition.
• To enhance the performance of models in high-resolution dense prediction.

Main Methods:

• Proposed Frequency-Aware Feature Fusion (FreqFusion) incorporating three novel components.
• Developed an Adaptive Low-Pass Filter (ALPF) generator to reduce intra-class inconsistency by attenuating high-frequency components within objects.
• Introduced an offset generator for refining inconsistent features and boundaries via resampling, and an Adaptive High-Pass Filter (AHPF) generator to restore lost high-frequency boundary details.

Main Results:

• FreqFusion effectively reduces intra-category inconsistency by preserving feature integrity during upsampling.
• The method significantly sharpens object boundaries, mitigating displacement issues caused by feature blurring.
• Comprehensive visualizations and quantitative analyses confirmed the improvements in feature consistency and boundary accuracy.

Conclusions:

• Frequency-Aware Feature Fusion (FreqFusion) offers a significant advancement over traditional feature fusion techniques.
• The proposed method demonstrably enhances feature consistency and boundary precision in dense image prediction.
• FreqFusion proves effective across various dense prediction tasks, highlighting its versatility and impact.