TILLvisION Tutorial

TILLvisION - Image filters

Physical processes tend to generate a certain amount of noise. Likewise, most images taken with cameras and frame grabbers are noisy. One effective method to reduce noise in images makes use of spatial filtering. These operations, also called convolutions, determine each pixel's value based on the pixel's neighborhood and a convolution kernel. Two main filter categories allow noise reduction: low-pass filters and rank filters.

Source Image

Grayscale

Our example is part of a fluorescence image of HELA cells. To emphasize the effect of edge detection filters, the background of the image is empty (=0).

	Pseudo Color Display Usually you would display your images in pseudocolor to improve visualization. Here we use TILLvisION's "Rainbow 2" palette, but the color corresponding zero is changed to gray (from black).

Reducing Noise

Low-pass spatial filters can reduce Gaussian random noise. These filters replace each pixel with a weighted sum of each pixel's neighborhood. As these filters remove noise, they have the side-effect of generally smoothing or blurring images and reducing edge information.

	Process / Filter / Lowpass (Average) The averaging filter can be selected from 3x3 up to 255x255 neighborhoods. All pixels included in the filters size are added up, the result is divided by the number of pixels.
	Process / Filter / Lowpass (Gauss) The Gauss filter can be selected from 3x3 up to 7x7 kernel size. It performs a weighted sum (center pixel = highest weight). The result is normalized by total kernel weight.
	Process / Filters / Median With a kernel size of 3x3 the resulting value is the median (number 5 out of the sorted list of 9). In contrast to lowpass filters, the median keeps edges and removes single pixel errors (like hot pixels) completely.

Improving Special Image Information

	Kernel -1 -1 -1 -1 8 -1 -1 -1 -1	Laplace (8 connected) The Laplace Filter weights the difference between the center pixel and its neighbors.
	Kernel -1 -1 -1 -1 9 -1 -1 -1 -1	Sharpen (8 connected) Based on the Laplace Filter, the Sharpen Filter includes/adds the original image (center pixels weight one above Laplace)
	Kernel -1 -2 -1 -0 -0 -1 -1 -2 -1	Sobel The Sobel filter enhances edges in all directions. It is implemented through two independent convolutions with the left kernel (once rotated by about 90). The results of each kernel are combined to form the final result.
	Kernel 1 1 1 0 0 0 -1 -1 -1	Prewitt The Prewitt filter performs a similar operation as the Sobel Filter (90 kernel rotate).

The filters described above are only a small part of TILLvisION's filter capabilities. Almost any common filter is available predefined. If you need a filter operation not included you can define your own filter kernels and structuring elements.