Image Enhancement: Principles and the Resolution Barrier

When a camera captures an image, the scene is projected onto a sensor grid. Each photosite on that grid integrates the light falling on it and produces a single value. The detail that falls between grid positions is averaged away and is gone permanently. This is the fundamental constraint, and it applies equally to a phone camera, a professional DSLR, and a surveillance sensor.

The Nyquist-Shannon sampling theorem states that a signal can be faithfully reconstructed only if the sampling rate is at least twice the highest frequency present. For a digital image, that means a feature must span at least two pixels to be recorded without aliasing. A 5-pixel face at 20 metres contains almost no usable identity information, and that is not a camera-quality problem or a software limitation. It is physics.

The lens introduces a second constraint independent of the sensor. Even with infinite pixels, an imperfect lens blurs the image by spreading each point of light across a small region. The size and shape of that spread is the point spread function. A defocused lens produces a circular disk PSF. Camera motion during exposure produces a linear streak. Both are recoverable if the PSF can be estimated accurately enough, which is the premise of deconvolution-based sharpening.

A surveillance camera recording a dimly lit car park may capture a face with sufficient spatial resolution, but the tonal range in the image is compressed into the dark end of the histogram. The information is there; it is just hard to see. Contrast enhancement addresses this problem. It is a legitimate, scientifically sound forensic operation when properly documented.

1. Global histogram equalisation
   Redistributes pixel values so that each brightness level has roughly equal frequency across the image. Effective when the useful detail occupies most of the frame, but tends to over-brighten localised bright areas such as light sources or windows.
2. CLAHE (Contrast Limited Adaptive HE)
   Divides the image into small rectangular tiles, equalises the contrast within each tile, then blends adjacent tiles with bilinear interpolation. The contrast amplification in each tile is clipped at a user-defined limit to suppress noise amplification. The result: shadow detail recovered without washing out highlights.
3. Gamma correction
   A non-linear brightness adjustment. Raising gamma above 1 brightens the midtones without clipping highlights as aggressively as linear brightening. Commonly used to compensate for cameras operating outside their intended exposure range.

Deconvolution treats the blurred image as the result of convolving the true scene with the PSF, plus noise. The goal is to estimate the true scene by inverting the convolution. Two broad families of algorithm exist in forensic practice.

The Wiener filter is the classical approach: it minimises mean squared error between the recovered image and the true scene given knowledge of both the PSF and the noise spectrum. In practice, the PSF is rarely known exactly, and any error in the PSF estimate produces characteristic ringing artefacts around edges. The analyst's obligation is to demonstrate that ringing has not been mistaken for real detail, particularly at forensically important locations such as a person's features or a vehicle identifier.

Super-resolution describes any method that outputs an image with more pixels than the source. The two fundamentally different approaches are frequently confused in forensic contexts, with important consequences for admissibility.

• Interpolation-based SR: bicubic, Lanczos, and related methods estimate new pixel values from their neighbours using a mathematical formula. They produce a smoother image with more grid positions, but they cannot add information that was not in the source. The result looks sharper at a glance but contains no new forensic data.
• Learning-based SR (deep neural networks): convolutional networks trained on large image datasets learn statistical relationships between low-resolution and high-resolution patches. Applied to a new low-resolution input, they hallucinate likely high-resolution content based on what they have seen before. The output can look strikingly sharper, but the added detail is generated by the model, not derived from the original capture.

Multi-frame super-resolution is a distinct, more defensible technique. When a surveillance camera records the same scene across multiple frames and the camera or subject moves slightly between frames, sub-pixel shifts create genuinely new sampling positions. Registering and combining those frames can legitimately recover detail up to the sensor's theoretical limit. The key difference: the information was present in the original recording, distributed across frames, not generated by a model.

The Scientific Working Group for Imaging Technologies produced guidelines that remain the practical standard in US courts and are widely adopted internationally. The core requirements are straightforward but must be followed completely.

1. Preserve the original: the unaltered source file must be secured, hashed (typically SHA-256), and retained as the primary exhibit before any enhancement work begins.
2. Document every step: each processing operation must be recorded in a processing log, naming the software, version, parameters, and purpose of the step.
3. Distinguish enhancement from interpretation: contrast and blur correction are enhancement. Identifying a person from the result is interpretation. The two must not be conflated in the report.
4. Disclose what was not enhanced: sections of the image not subjected to any processing must also be identified, so the court knows the scope of the analyst's work.
5. Present both: in court, the original and the enhanced image should be shown together. The enhancement is a tool for perception, not a substitute for the evidence.

Forensic image enhancement generates two categories of output: positive conclusions supported by the enhanced detail (this appears to be a dark-coloured saloon car) and negative conclusions about what cannot be determined (insufficient resolution to confirm or exclude a specific vehicle registration). Both must be stated explicitly in the report.

Processed images carry high perceived credibility with courts and juries, often exceeding their actual information content. Noise reduction and contrast stretching can make a face feel recognisable even when pixel count is unchanged. The forensic analyst must counter this by reporting the pixel dimensions of the original face region, the estimated spatial resolution in line pairs per millimetre, and an explicit statement of what level of identification the data actually supports.