Abstract
As a 3-order tensor, a multi-spectral image (MSI) has dozens of spectral bands, which can deliver more information for real scenes. However, real MSIs are often corrupted by noises in the sensing process, which will further deteriorate the performance of higher-level classification and recognition tasks. In this paper, we propose a Low-rank Tensor Dictionary Learning (LTDL) method for MSI denoising. Firstly, we extract blocks from the MSI and cluster them into groups. Then instead of using the exactly low-rank model, we consider a nearly low-rank approximation, which is closer to the latent low-rank structure of the clean groups of real MSIs. In addition, we propose to learn a spatial dictionary and a spectral dictionary, which contain the spatial features and spectral features respectively of the whole MSI and are shared among different groups. Hence the LTDL method utilizes both the latent low-rank prior of each group and the correlation of different groups via the shared dictionaries. Experiments on synthetic data validate the effectiveness of dictionary learning by the LTDL. Experiments on real MSIs demonstrate the superior denoising performance of the proposed method in comparison to state-of-the-art methods.
Introduction
Multi-spectral images (MSIs) capture dozens of spectral bands, providing significantly more information than standard RGB images. However, the sensing process often introduces noise that degrades performance in critical tasks like classification and object recognition. The primary challenge in denoising MSIs lies in balancing spatial correlation (patterns within a single band) and spectral correlation (patterns across different bands).
Traditional methods often fail by either treating bands independently or enforcing an "exactly" low-rank constraint. A critical insight of this work is that real-world MSIs are not perfectly low-rank; they possess a "long tail" of small singular values that contain essential texture information. Forcing these to zero results in blurry, over-smoothed images.
Methodology
The proposed Low-rank Tensor Dictionary Learning (LTDL) method introduces two key innovations to address the limitations of existing models:
- Shared Dictionaries: Instead of learning separate dictionaries for different parts of an image, LTDL learns a spatial dictionary ($D^a$) and a spectral dictionary ($D^e$) shared across all groups of similar image blocks. This "multitask learning" approach captures global features more effectively by recognizing that different clusters (e.g., grass vs. buildings) share underlying spatial and spectral "atoms."
- Nearly Low-rank Approximation: Rather than enforcing a hard rank constraint, the model allows for a "nearly" low-rank structure. It decomposes each group of image blocks into a low-rank component and a residual component, ensuring that the texture-rich "long tail" of singular values is preserved.
The optimization problem is solved using the Alternating Direction Method of Multipliers (ADMM), which breaks the complex non-convex problem into manageable sub-problems updated iteratively. By treating the MSI as a 3rd-order tensor throughout, the method avoids the pitfalls of data "unfolding" that can destroy inherent multi-dimensional structures.
Experimental Results
The researchers evaluated LTDL using synthetic data, the Columbia benchmark dataset (32 real-world scenes), and HYDICE MSI satellite data. LTDL consistently outperformed state-of-the-art methods like K-SVD, BM3D, and KBRreg across multiple metrics:
- Quantitative Performance: On the Columbia dataset ($\nu=0.1$), LTDL achieved a PSNR of 41.50, significantly beating standard BM3D (36.92) and the next best method, KBRreg (40.63).
- Spectral Fidelity: It achieved the lowest Spectral Angle Mapper (SAM) scores, indicating superior preservation of the pixels' spectral signatures.
- Qualitative Quality: Visual results demonstrate that LTDL preserves fine details, such as the centers of flowers, which other methods often blur into uniform masses. It also effectively removed complex "stripe noise" in real-world satellite imagery.
Conclusion
LTDL advances MSI denoising by moving away from rigid low-rank assumptions. By utilizing shared dictionaries and nearly low-rank approximations, it preserves fine textures while effectively suppressing noise. Future work will focus on formal convergence proofs and computational optimizations, such as GPU acceleration, to handle high-dimensional tensors more efficiently for real-time applications.
Summary updated