Image processing

Most one- or two-dimensional signals in everyday life (audible or visible) arise in analogue form. Optical imaging systems or electrical networks process information in an analogue way. They are excellent in execution speed; however, complicated analogue signal or image processing algorithms are very difficult and sometimes impossible to implement.

Today, most signals are converted into a form tractable by digital hardware, and can then be treated by digital signal processing (for one-dimensional signals) or by image processing (in two dimensions).

To convert analogue signals into a digital form one has to sample them sufficiently frequently ( Sampling Theorem) and quantize (digitize) the samples, a process usually called a nalogue-to-digital conversion (ADC). After digital processing (DSP), the digital signal is sometimes converted back to an analogue form (DAC) as, e.g. in an image processing system (for viewing). In this case, a video signal from a television camera is the analogue input, the processing is done by some digital hardware in or close to real time (possibly by specialized processors), and the output is again a video signal.

Transmission and/or storage of digital signals sometimes needs as a first step, for reasons of economy, data compression. Normally the signals have to be improved in some sense ( Filtering, Image Enhancement, Sharpening, Smoothing). Enhancement with the aim of getting rid of some degradation known a priori is called image restoration.

One of the most important parts of practically any automated image recognition system is called image segmentation. This is the classification of each image pixel into one of the constituent image parts.

Signal or image processing methods can be executed either directly in the time or spatial domain, respectively, or one can first transform the signals into another domain ( Orthogonal Functions), perform the processing in the transform domain, and then perform the back transformation. Transformations of some input functions f(x,y,z;t) into some output functions g(x,y,z;t) can often be treated as linear shift-invariant systems, for which convolution is the standard operation. For many enhancement problems, non-linear methods like rank filters or morphological operations are indicated.

Digital signal or image processing has found many application in today's commodity markets, and can be extremely compute-intensive. Much effort went into the methods, but also into the development of fast algorithms and into computer architectures ( Parallel Processing).

Here is a choice of standard textbooks on signal processing: Kunt80, Rabiner75, Oppenheim75; on image processing: Jain89, Gonzalez87, Pratt78, Rosenfeld76; on specialized hardware:Kung88.