3-D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising

Area of Science:

• Computer Vision
• Machine Learning
• Signal Processing

Background:

• Hyperspectral images (HSIs) are susceptible to noise, degrading their quality and utility.
• Existing denoising methods struggle to effectively capture complex spatiospectral correlations within HSIs.

Purpose of the Study:

• To propose an advanced neural network for hyperspectral image denoising.
• To effectively model both structural spatiospectral correlation and global correlation along the spectrum (GCS).

Main Methods:

• Developed an alternating directional 3-D quasi-recurrent neural network (QRNN3D).
• Utilized 3-D convolution for extracting structural spatiospectral correlation.
• Employed a quasi-recurrent pooling function to capture GCS.
• Introduced an alternating directional structure to eliminate causal dependence without added computational cost.

Main Results:

• The proposed model effectively models spatiospectral dependence in HSIs.
• Demonstrated significant improvements in restoration accuracy compared to state-of-the-art methods.
• Achieved reduced computation time for HSI denoising.
• The model shows flexibility for HSIs with varying numbers of bands.

Conclusions:

• The alternating directional 3-D quasi-recurrent neural network offers superior performance for HSI denoising.
• This approach effectively balances restoration accuracy and computational efficiency.
• The method is adaptable to HSIs with diverse spectral characteristics.