Decide to do

I’m starting with the Sobel filter for the nth time, trying to understand it from a different perspective.

Basically, it detects changes in the x and y directions, then combines them using the hypotenuse formula. It’s like taking a vector in the fourth quadrant. The filter also uses some collective reference of nearby pixels — something like an average but with a different formula.

If we “zoom in” on the pixel values, it essentially reveals gaps in the intensities. In short, it detects changes, magnifies them, and stores that information as a value.

Now, does this value remain unique? It’s like a vector in the fourth quadrant, where the magnitude and direction depend on the surrounding pixels. So yes, it can be unique — unless the distribution changes. When the distribution changes, the direction changes too, and so we end up with different values.

You can think of it like an image represented by a magnitude and direction in space. For larger images, the filter’s magnitude can represent more complex or higher-dimensional patterns.

This raises a critical question: How do we choose the filter size, and how many times should we apply it?

• For small details, go with a smaller filter size.

• For detecting larger features, use a larger filter size.

Honestly, that’s quite a conclusion — not gonna lie.

Interestingly, the edge cases in the Sobel filter could inspire another algorithm, like Canny edge detection, where you set thresholds for different lighting conditions — a bit like choosing different varieties of coffee.

In Canny, the filter’s “center” is placed in the middle of the window, whereas in Sobel, we start computing gradients from the first pixel.

Here’s a proposal: What if we make things harder and first convert the image grid into a circular plane, then find the diameter? Sounds nice, not gonna lie.