Abstract
Hyperspectral image (HSI) denoising is a critical task in remote sensing, as noise significantly degrades image visibility and analysis accuracy. Current methods often struggle to balance the extraction of global context with the preservation of local spatial textures. This paper introduces PSFormer, a novel denoising model that integrates Transformers with parallel channel-shuffling convolutional neural networks (CNNs). By employing a Channel Shuffling Module (CSM), a Block Downsampling Global Enhancement Module (DGEM), and an Adaptive Bidirectional Feature Fusion Module (ABFM), the proposed architecture effectively captures both long-range dependencies and fine-grained details. Experimental results on the Pavia and ICVL datasets demonstrate that PSFormer significantly outperforms existing state-of-the-art methods, achieving up to a 2.18 dB improvement in Peak Signal-to-Noise Ratio (PSNR) under complex noise conditions.
Introduction
Hyperspectral images provide rich spectral information across hundreds of bands, but they are highly susceptible to noise from sensors and atmospheric interference. Traditional CNN-based denoisers are limited by local receptive fields, failing to capture the global spatial patterns necessary for high-quality restoration. Conversely, while Transformers excel at global context modeling, their structural differences from CNNs make seamless integration challenging. PSFormer addresses these issues through a dual-branch architecture that processes local and global features in parallel, ensuring comprehensive feature extraction and robust denoising.
Methodology
The PSFormer architecture is built around three core components designed to optimize information flow and feature integration:
- Channel Shuffling Module (CSM): This module handles the high dimensionality of HSI data. By grouping and shuffling channels, it promotes information exchange across the spectral dimension while extracting local spatial features efficiently without a massive increase in parameters.
- Block Downsampling Global Enhancement Module (DGEM): To capture global structures while maintaining manageable computational costs, this Transformer-based branch uses a block downsampling strategy. This increases the receptive field, allowing the model to perceive larger image regions.
- Adaptive Bidirectional Feature Fusion Module (ABFM): This module acts as a bridge, consisting of Global Feature Fusion (GFM) and Local Feature Fusion (LFM) components. It allows for two-way interaction, ensuring that the local branch benefits from global context and the global branch retains fine-grained spatial details.
Multi-Scale Channel Mixing
A key novelty is the CSM's ability to perform multi-scale channel mixing. As shown in the detailed module diagram, the system uses group convolutions and shuffling to ensure that features from different spectral bands are effectively integrated, which is crucial for maintaining spectral integrity during the denoising process.
Experimental Results
The model was rigorously tested against five mainstream denoising methods, including BM4D, LRMR, HSID-CNN, and QRNN3D, using the ICVL and Pavia University datasets.
- Quantitative Performance: On the Pavia dataset under high Gaussian noise, PSFormer improved PSNR by up to 0.4 dB over the next best model. On the ICVL dataset with mixed noise (Gaussian, impulse, and stripes), the improvement reached a substantial 2.18 dB.
- Qualitative Quality: Visual comparisons indicate that PSFormer excels at recovering sharp edges and intricate textures where other models produce over-smoothed or artifact-laden results.
- Spectral Fidelity: Using the Spectral Angle Mapper (SAM) metric, the researchers confirmed that PSFormer preserves the unique spectral signatures of materials better than competing models, which is vital for downstream classification tasks.
Conclusion
PSFormer represents a significant advancement in HSI denoising by successfully merging the strengths of CNNs and Transformers. The parallel architecture and adaptive fusion mechanism allow for superior noise removal while preserving the rich spatial and spectral information inherent in hyperspectral data. Future work will focus on optimizing the model's computational efficiency for real-time remote sensing applications and exploring its utility in other tasks like HSI super-resolution.
Summary updated