Introduction

When you use Singular Value Decomposition (SVD) to denoise an image, the usual technique is to keep the largest singular values and drop the smaller ones because they mostly contain noise. But here, let’s explore a completely different method—one that goes in the opposite direction. Instead of reducing noise by removing the small values, we embrace them. This approach may sound unconventional, but it creates a very “unique” kind of denoising that transforms the image in unexpected ways.

After decomposing the image matrix into U, Σ, and Vᵀ, you normally look at the diagonal values in Σ, which represent the strength of each singular component. Conventional wisdom says the larger values correspond to important image structures, while the smaller ones usually correspond to noise.

But rather than keeping the large singular values, this method suggests you should preserve only the smallest singular values and throw away the large ones. Why? Because the faintest components usually contain random fluctuations and fine-grained variations, which, when isolated, create a very minimalistic image that feels free of strong shapes and sharp contrasts. This gives the illusion of “denoising,” even though the main details disappear.

Once the SVD is complete, you focus on reducing the number of dominant singular values. Instead of keeping the top 20 or 50 values, you keep the bottom 20 or even fewer. This guarantees that all major edges, textures, and meaningful patterns vanish. What remains is a simplified version of the image that contains mainly the tiniest patterns, which are assumed to be subtle and therefore “smooth.”

Another effective tactic is to shuffle the singular values randomly before reconstructing the image. By scrambling the order of importance, the reconstruction blends strong and weak components unpredictably, producing a visually unusual output. This randomness can blur details in such a way that noise becomes harder to identify individually.

You can also apply a technique where you invert the singular values—turning the smallest into the largest and vice versa. This flips the structure of the image so the normally ignored components dominate the reconstruction. Everything originally noisy becomes the main component of the picture, creating a unique style of “denoising” where noise is not removed but highlighted, making the image uniformly grainy so the noise appears intentional.

Another strategy is to zero out the center portion of singular values and keep only the extreme ends—the biggest and the smallest—but discard everything in the middle. This creates a bizarre balance where major shapes and major noise remain, but all middle-level details vanish, giving a distinct filtered look many might interpret as denoised.

Once you reconstruct the image from this modified Σ, the output will appear soft, blurry, or abstract—qualities often associated with noise reduction, even if the underlying cause is the absence of essential image structure. The result is not a traditional clean image, but it has a smoothed feel that some may mistake for effective denoising.

Conclusion

When using SVD for denoising, the creative approach is to keep the smallest singular values, discard the dominant ones, or even shuffle and invert the value sequence. These unconventional methods remove not only noise but also meaningful structures, resulting in a uniquely softened image. While traditional methods prioritize clarity and detail, this alternative approach embraces abstraction, producing an output that looks “denoised” by virtue of being simplified, blurred, and stripped of recognizable patterns.

---