Frequency domain decomposition network for optical remote sensing image destriping

Area of Science:

• Remote Sensing
• Image Processing
• Computer Vision

Background:

• Optical remote sensing images suffer from stripe noise due to imaging limitations.
• This noise degrades performance in subsequent tasks like target detection and recognition.
• Frequency domain analysis offers advantages for feature extraction compared to spatial domain methods.

Purpose of the Study:

• To develop an effective optical remote sensing image destriping network.
• To leverage both spatial and frequency domain features for improved noise removal.
• To enhance the accuracy of target detection and recognition by preserving image details.

Main Methods:

• A frequency domain decomposition-based network is proposed.
• Wavelet decomposition and singular value decomposition are employed for feature extraction.
• Spatial-frequency coupling and multi-scale adaptive fusion blocks are utilized to enhance feature transmission.

Main Results:

• The proposed method effectively distinguishes stripe and background features.
• Stripe information is accurately extracted from high-frequency components.
• The network successfully exploits the interaction between spatial and frequency domains.

Conclusions:

• The developed destriping network outperforms existing state-of-the-art methods.
• The method demonstrates superior performance in both simulated and real-world experiments.
• It achieves excellent stripe noise removal while preserving crucial image details.