Marr-Hildreth(高斯拉普拉斯)滤波器的复杂度、缺陷及选用高斯差分的原因

Hey there, let's break these two questions down clearly—they’re fundamental when working with edge detection and scale-space computer vision methods!

The Marr-Hildreth filter operates in two core steps: first applying Gaussian smoothing to the image, then computing the Laplacian of the smoothed result. Here's how the complexity breaks down:

• Naive implementation: If you use a non-separable k×k Gaussian kernel on an N×N image, the convolution step runs at O(N²k²). Each of the N² pixels requires k² multiply-add operations, which gets costly as the kernel size grows.
• Optimized (separable) implementation: Gaussian kernels are separable—you can split a k×k kernel into a 1×k horizontal kernel and a k×1 vertical kernel. First convolve the image with the horizontal kernel (O(N²k)), then the vertical one (another O(N²k)). Total complexity drops to O(N²k), which is far more efficient for larger kernels.
• The Laplacian step typically uses a small fixed kernel (like 3×3), so its complexity is O(N²)—negligible compared to the Gaussian smoothing cost.

Overall, the complexity is dominated by how you implement the Gaussian convolution; the separable version is the standard in practical applications.

Let’s start with the flaws of LoG, then jump to why DoG is a better choice for most tasks:

Key Defects of LoG #

• Double-edge response: The LoG’s zero-crossing detection often generates a pair of positive/negative responses on either side of a real edge. This creates ambiguity in edge localization, requiring extra post-processing to clean up or merge these duplicate signals.
• Redundant computation: Even with separable kernels, computing LoG requires first smoothing with Gaussian, then applying the Laplacian. For multi-scale tasks (where you analyze the image at different σ values), this means re-computing Gaussian filters and Laplacians for each scale—wasting unnecessary cycles.
• Noise amplification: While Gaussian smoothing reduces noise, the Laplacian’s second derivative still amplifies high-frequency noise more noticeably than DoG, especially in fine-grained scales.

Why DoG is the Preferred Alternative #

• Massive computational efficiency: DoG is calculated by subtracting two adjacent layers in a Gaussian pyramid. If you’re already building a Gaussian pyramid for multi-scale analysis, DoG comes essentially for free—no extra convolution steps needed. This cuts down computation time drastically compared to computing LoG for each scale.
• Cleaner edge responses: DoG doesn’t produce the strong double-edge artifact that LoG does, making edge or keypoint detection more straightforward without extra cleanup steps.
• Accurate approximation of LoG: Mathematically, DoG is a very close approximation of LoG when the scale ratio between adjacent Gaussian layers is ~1.6 (σ₂ = 1.6σ₁). This approximation is more than accurate enough for real-world computer vision tasks (like SIFT feature detection), so you get nearly identical scale-space properties without the LoG’s drawbacks.

内容的提问来源于stack exchange，提问作者van