Image Processing, Multimedia and Biomedical Imaging

An image is a two-dimensional signal, and most of signal processing carries over to it directly. What makes imaging a discipline of its own is that the second dimension introduces genuinely new structure — edges, textures, regions, geometry — and that in many of the most important applications the picture is not captured at all. It is computed from measurements that look nothing like a picture.

What a Digital Image Actually Is

A digital image is a grid of samples. Each pixel records intensity, usually across several colour channels. Two parameters set the ceiling on what can be recovered from it.

Spatial resolution is how finely the scene is sampled. It is subject to the same sampling theorem as any other signal, and violating it produces the two-dimensional form of aliasing: the shimmering false patterns visible on finely striped fabric in photographs and video. As with one-dimensional signals, this must be prevented optically or at the sensor — it cannot be undone afterwards.

Bit depth is how finely intensity is quantised. Eight bits per channel is adequate for display but often inadequate for analysis, because it discards subtle gradations that matter when an image is measured rather than merely looked at. Scientific and medical imaging routinely uses twelve, fourteen or sixteen bits for exactly this reason.

The Basic Operations

Point operations transform each pixel independently — brightness, contrast, gamma correction. Cheap, and unable to use any spatial context.

Spatial filtering combines each pixel with its neighbours. Smoothing filters average locally, suppressing noise at the cost of blurring genuine detail. Sharpening filters do the opposite, enhancing detail while amplifying noise. The tension is fundamental: at the pixel level, fine detail and noise are the same thing — rapid local variation — and no filter can separate them without additional assumptions about which is which.

Edge detection looks for locations where intensity changes rapidly, on the reasoning that such changes usually mark boundaries between objects. It is a gradient computation, which means it is a differentiation, which means it amplifies noise — so edge detectors invariably smooth first. The choice of how much to smooth determines whether fine real edges are missed or noise is reported as structure.

Segmentation partitions an image into meaningful regions. This is where straightforward processing ends and interpretation begins, because "meaningful" depends entirely on the question being asked. The same image segmented for one purpose is segmented wrongly for another.

Compression: Discarding What Is Not Noticed

Raw images are large, and compression exploits two distinct kinds of redundancy.

Lossless compression removes statistical redundancy — neighbouring pixels are usually similar, so the differences between them compress well. The original is recovered exactly. Modest ratios, no compromise.

Lossy compression additionally discards information judged perceptually unimportant. It transforms blocks of the image into a frequency representation, then quantises the high-frequency coefficients coarsely, exploiting the fact that human vision is less sensitive to fine high-frequency detail than to broad structure and to luminance more than colour.

The ratios achieved are far larger, which is why lossy formats dominate consumer imaging. But the discarded information is gone, and the characteristic artefacts — blocking, ringing near sharp edges, colour bleeding — become severe under repeated re-encoding. For any image that will be measured rather than viewed, lossy compression is a serious hazard: it alters exactly the fine detail that analysis depends on, and it does so invisibly to casual inspection.

Biomedical Imaging: Pictures That Are Reconstructed, Not Captured

Medical imaging is where the field becomes most interesting, because in most modalities no camera ever sees the subject. Raw measurements bear no resemblance to an image, and the picture exists only after solving a mathematical inverse problem. The NIBIB science topics collection is a reliable plain-language introduction to the major methods.

Computed tomography

CT measures X-ray attenuation along many straight paths through the body from many angles. Each measurement is a single number: total absorption along one line. Reconstruction infers the three-dimensional distribution of attenuation consistent with all of those line integrals — a classic inverse problem, and the reason CT arrived only once sufficient computing power existed.

Magnetic resonance imaging

MRI places the body in a strong magnetic field, excites hydrogen nuclei with radio-frequency pulses, and detects the signal they emit as they relax. Spatial information is encoded by deliberately varying the magnetic field across space, so that position maps to frequency and phase. The raw data are therefore acquired directly in a spatial-frequency domain, and the image is produced by an inverse Fourier transform — one of the most direct practical applications of the frequency-domain view described on the signal processing page. The magnetic resonance imaging overview covers the physics accessibly.

Ultrasound

Ultrasound transmits acoustic pulses and times the echoes returning from tissue boundaries. Because it uses sound rather than ionising radiation and produces images in real time, it occupies a distinct clinical niche; its resolution and penetration trade against one another through the choice of frequency, since higher frequencies resolve finer detail but attenuate faster with depth. See the ultrasound imaging overview.

Why Reconstruction Artefacts Matter

Because these images are computed, they can contain structures that are artefacts of the reconstruction rather than features of the subject: streaks radiating from dense objects in CT, geometric distortion from field inhomogeneity in MRI, acoustic shadowing in ultrasound.

These artefacts look like image content because they are image content — the algorithm produced them. Recognising them is a trained skill and one of the reasons imaging expertise cannot be reduced to running software. It also explains why calibration and quality assurance are as central to imaging as the reconstruction mathematics itself.