Computer Vision
Image Filters
A filter is a small square of numbers. Slide it over a photograph, multiply and add, and the same handful of weights will blur it, sharpen it or leave nothing but its edges. Pick one, resize it, or drag the weights until it is yours.
a positive weighta negative weight
Drag a weight up or down to change it · the picture follows immediately
Filter—
—
Photo
Channels
Size
Filter
the weights add to—
multiplications a pixel—
output before clipping—
What to observe
- Weights that add to one keep the average brightness, so a sharpened photograph still looks like the photograph.
- Weights that add to zero keep only what changes: flat areas return nothing and outline leaves the edges on black.
- Sobel X is left minus right and lights up vertical edges; Sobel Y is the same weights turned on their side.
- A random filter is almost always noise. A convolutional network finds the useful ones by gradient descent; where the window may sit is Convolution.
Shortcuts:f fullscreen