Frequency Domain

For simplicity, assume that the image I being considered is formed by projection from scene S (which might be a two- or three-dimensional scene, etc.).
The frequency domain is a space in which each image value at image position F represents the amount that the intensity values in image I vary over a specific distance related to F. In the frequency domain, changes in image position correspond to changes in the spatial frequency, (or the rate at which image intensity values) are changing in the spatial domain image I.
For example, suppose that there is the value 20 at the point that represents the frequency 0.1 (or 1 period every 10 pixels). This means that in the corresponding spatial domain image I the intensity values vary from dark to light and back to dark over a distance of 10 pixels, and that the contrast between the lightest and darkest is 40 gray levels (2 times 20).
The spatial frequency domain is interesting because: 1) it may make explicit periodic relationships in the spatial domain, and 2) some image processing operators are more efficient or indeed only practical when applied in the frequency domain.
In most cases, the Fourier Transform is used to convert images from the spatial domain into the frequency domain and vice-versa.
A related term used in this context is spatial frequency, which refers to the (inverse of the) periodicity with which the image intensity values change. Image features with high spatial frequency (such as edges) are those that change greatly in intensity over short image distances.

Color Quantization

Color quantization is applied when the color information of an image is to be reduced. The most common case is when a 24-bit color image is transformed into an 8-bit color image.
Two decisions have to be made:
  1. which colors of the larger color set remain in the new image, and
  2. how are the discarded colors mapped to the remaining ones.

The simplest way to transform a 24-bit color image into 8 bits is to assign 3 bits to red and green and 2 bits to blue (blue has only 2 bits, because of the eye's lower sensitivity to this color). This enables us to display 8 different shades of red and green and 4 of blue. However, this method can yield only poor results. For example, an image might contain different shades of blue which are all clustered around a certain value such that only one shade of blue is used in the 8-bit image and the remaining three blues are not used.
Alternatively, since 8-bit color images are displayed using a colormap, we can assign any arbitrary color to each of the 256 8-bit values and we can define a separate colormap for each image. This enables us perform a color quantization adjusted to the data contained in the image. One common approach is the popularity algorithm, which creates a histogram of all colors and retains the 256 most frequent ones. Another approach, known as the median-cut algorithm, yields even better results but also needs more computation time. This technique recursively fits a box around all colors used in the RGB colorspace which it splits at the median value of its longest side. The algorithm stops after 255 recursions. All colors in one box are mapped to the centroid of this box.
All above techniques restrict the number of displayed colors to 256. A technique of achieving additional colors is to apply a variation of half-toning used for gray scale images, thus increasing the color resolution at the cost of spatial resolution. The 256 values of the colormap are divided into four sections containing 64 different values of red, green, blue and white. As can be seen in Figure 1, a 2×2 pixel area is grouped together to represent one composite color, each of the four pixels displays either one of the primary colors or white. In this way, the number of possible colors is increased from 256 to Eqn:eqnquant.

Figure 1 A 2×2 pixel area displaying one composite color.


Convolution is a simple mathematical operation which is fundamental to many common image processing operators. Convolution provides a way of `multiplying together' two arrays of numbers, generally of different sizes, but of the same dimensionality, to produce a third array of numbers of the same dimensionality. This can be used in image processing to implement operators whose output pixel values are simple linear combinations of certain input pixel values.
In an image processing context, one of the input arrays is normally just a graylevel image. The second array is usually much smaller, and is also two-dimensional (although it may be just a single pixel thick), and is known as the kernel. Figure 1 shows an example image and kernel that we will use to illustrate convolution.

Figure 1 An example small image (left) and kernel (right) to illustrate convolution. The labels within each grid square are used to identify each square.

The convolution is performed by sliding the kernel over the image, generally starting at the top left corner, so as to move the kernel through all the positions where the kernel fits entirely within the boundaries of the image. (Note that implementations differ in what they do at the edges of images, as explained below.) Each kernel position corresponds to a single output pixel, the value of which is calculated by multiplying together the kernel value and the underlying image pixel value for each of the cells in the kernel, and then adding all these numbers together.
So, in our example, the value of the bottom right pixel in the output image will be given by:

If the image has M rows and N columns, and the kernel has m rows and n columns, then the size of the output image will have M - m + 1 rows, and N - n + 1 columns.
Mathematically we can write the convolution as:

where i runs from 1 to M - m + 1 and j runs from 1 to N - n + 1.
Note that many implementations of convolution produce a larger output image than this because they relax the constraint that the kernel can only be moved to positions where it fits entirely within the image. Instead, these implementations typically slide the kernel to all positions where just the top left corner of the kernel is within the image. Therefore the kernel `overlaps' the image on the bottom and right edges. One advantage of this approach is that the output image is the same size as the input image. Unfortunately, in order to calculate the output pixel values for the bottom and right edges of the image, it is necessary to invent input pixel values for places where the kernel extends off the end of the image. Typically pixel values of zero are chosen for regions outside the true image, but this can often distort the output image at these places. Therefore in general if you are using a convolution implementation that does this, it is better to clip the image to remove these spurious regions. Removing n - 1 pixels from the right hand side and m - 1 pixels from the bottom will fix things.
Convolution can be used to implement many different operators, particularly spatial filters and feature detectors. Examples include Gaussian smoothing and the Sobel edge detector .

Multi-spectral Images

A multi-spectral image is a collection of several monochrome images of the same scene, each of them taken with a different sensor. Each image is referred to as a band. A well known multi-spectral (or multi-band image) is a RGB color image, consisting of a red, a green and a blue image, each of them taken with a sensor sensitive to a different wavelength. In image processing, multi-spectral images are most commonly used for Remote Sensing applications. Satellites usually take several images from frequency bands in the visual and non-visual range. Landsat 5, for example, produces 7 band images with the wavelength of the bands being between 450 and 1250 nm.
All the standard single-band image processing operators can also be applied to multi-spectral images by processing each band separately. For example, a multi-spectral image can be edge detected by finding the edges in each band and than ORing the three edge images together. However, we would obtain more reliable edges, if we associate a pixel with an edge based on its properties in all three bands and not only in one.
To fully exploit the additional information which is contained in the multiple bands, we should consider the images as one multi-spectral image rather than as a set of monochrome graylevel images. For an image with k bands, we can then describe the brightness of each pixel as a point in a k-dimensional space represented by a vector of length k.
Special techniques exist to process multi-spectral images. For example, to classify a pixel as belonging to one particular region, its intensities in the different bands are said to form a feature vector describing its location in the k-dimensional feature space. The simplest way to define a class is to choose a upper and lower threshold for each band, thus producing a k-dimensional `hyper-cube' in the feature space. Only if the feature vector of a pixel points to a location within this cube, is the pixel classified as belonging to this class. A more sophisticated classification method is described in the corresponding worksheet.
The disadvantage of multi-spectral images is that, since we have to process additional data, the required computation time and memory increase significantly. However, since the speed of the hardware will increase and the costs for memory will decrease in the future, it can be expected that multi-spectral images will become more important in many fields of computer vision.