Abstract: Based on the Total Variation (TV) norm minimization model, this paper introduces a novel image reconstruction algorithm, CG-TV. By leveraging fusion information within the undersampling domain, a compressed sensing-based image fusion model is also proposed. Numerical experiments demonstrate that the CG-TV algorithm reduces the required sampling rate compared to traditional methods while maintaining high reconstruction quality. Furthermore, the proposed fusion model exhibits superior performance across various image categories, including medical and multi-focus datasets.
Introduction
Traditional digital imaging relies on the Nyquist-Shannon sampling theorem, which requires sampling at twice the highest frequency of the signal. This leads to massive data volumes, straining storage and transmission hardware. Compressed Sensing (CS) offers a way out: if a signal is "sparse" in some domain, it can be reconstructed from far fewer samples than Nyquist requires. While many images aren't sparse in the pixel domain, their gradients (differences between adjacent pixels) usually are—a property exploited by Total Variation (TV) models. This paper addresses the challenge of developing an efficient, stable algorithm to solve the TV minimization problem and applying it to image fusion.
The CG-TV Algorithm
The authors introduce the CG-TV algorithm, which solves the optimization problem: $$\min \|x\|_{TV} \text{ subject to } y = \Phi x$$ where $x$ is the image, $y$ is the undersampled data, and $\Phi$ is the measurement matrix. Key technical components include:
- Conjugate Gradient (CG) Core: Unlike standard gradient descent, CG finds the optimal reconstruction path more efficiently, reducing the number of iterations needed.
- Iterative Refinement: Each step involves two critical sub-processes: iterative projection to ensure the estimate matches sampled frequency data, and TV denoising to suppress noise while preserving sharp edges.
- CS-Based Fusion Model: Instead of reconstructing two images and then fusing them, the authors propose fusing the undersampled measurements ($y_1$ and $y_2$) first, then performing a single reconstruction. This reduces the computational load at the receiver end.
Experimental Results
The study evaluates the algorithm using standard benchmarks (Lena, Cameraman) and specialized medical imaging. On an angiogram test, CG-TV achieved a PSNR of ~35.01 dB, a massive improvement over Wavelet-OMP, which sat at ~26.08 dB for the same sampling rate.
One of the most significant advantages of the TV-based approach is the preservation of structural integrity. While wavelet-based methods often struggle with "ringing" artifacts at low sampling rates, CG-TV maintains sharp edges and connectivity in high-contrast regions, such as blood vessels.
Key Insights and Conclusion
- Efficiency of CG: By using the Conjugate Gradient method for the $L_1$ norm approximation, the algorithm achieves faster convergence and higher PSNR than standard iterative methods.
- Fusion in the "Compressed" Domain: Fusing in the measurement domain maintains higher structural similarity (SSIM) because the reconstruction process itself acts as a regularizer, smoothing out inconsistencies between source images.
While the current model is highly effective, future work may involve integrating multi-scale analysis (like Curvelets or Shearlets) to further improve the reconstruction of complex textures and reducing the computational weight for real-time applications on low-power edge devices.
Summary updated