FSSM: Frequency-Enhanced State Space Modeling with FFT-Based Two-Sided Non-Causal Convolution for Image Dehazing

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Image dehazing is crucial for visual perception in adverse weather, particularly for UAV remote sensing, traffic monitoring, and surveillance.
• Existing methods like CNNs struggle with long-range dependencies, while Transformers are computationally expensive.

Purpose of the Study:

• To propose an efficient image dehazing framework, FSSM, that overcomes limitations of existing methods.
• To enhance both global dependency modeling and local texture details for superior image restoration.

Main Methods:

• Developed FSSM, integrating frequency-enhanced State Space Modeling with a hierarchical encoder-decoder architecture.
• Introduced FFT-based State Space Block (FFTSSB) for efficient bidirectional global dependency modeling via frequency-domain convolution.
• Designed Frequency-Aware Discriminative Enhancement Block (FDEB) for local texture enhancement using spatial gating and frequency modulation.
• Constructed Frequency-Aware State Interaction (FASI) block to couple global and local enhancement stages.

Main Results:

• FSSM demonstrated superior restoration accuracy, structural consistency, and perceptual quality on the HazyDet dataset compared to existing methods.
• Ablation studies confirmed the effectiveness of the proposed FFT-based state modeling, frequency-aware enhancement, and hierarchical design.

Conclusions:

• FSSM offers an effective and efficient solution for image dehazing, balancing global context and local detail restoration.
• The proposed frequency-domain approach and hierarchical structure significantly contribute to improved visual restoration performance.