One of the most characteristic analyses in digital imaging, whether in the field of forensic image analysis, quality control in digitization tasks, or even printing, is that which occurs through noise.
One of the most characteristic forms of analysis through noise is through the signal-to-noise ratio (Signal to Noise Ratio, SNR); however, there are other tools or strategies such as noise analysis through the frequency spectrum, which shows the distribution of noise in the frequency domain, usually through the use of the Fourier Transform.
The frequency spectrum of noise, also called the noise power spectrum density (Power Spectrum Density), shows the distribution of noise in terms of density in relation to its frequencies. That is, the frequency of noise occurrence in the image will have more or less presence depending on its spatial frequency. Similarly, in the SFR-MTF, frequency is the strategy to quantify the capacity to reproduce detail; those artifacts that exceed certain frequencies will be barely perceptible to our visual system. Therefore, the representation of noise through its frequency spectrum has an important correlation with the appearance of noise.
Thus, an image can be represented through the spatial domain, that is, in its conventional x and y coordinates, or in the frequency domain. In both domains, we can discern different aspects of noise.
Noise of a random nature, which follows models such as Gaussian, can be studied through histogram representations, which organize the image information by frequency of occurrence.
On the other hand, noise of a periodic nature, that is, noise that forms repetitive patterns throughout the image, can be identified and studied in the frequency domain through this spectrum.
Understanding the frequency spectrum
The frequency spectrum describes the density (Y-axis) of occurrence of certain frequencies (X-axis) in the image; in this way, it is possible to describe the presence of noise in a specific area of the image based on its frequency.
As an example, one can start from a synthetic image with a pattern of lines separated by 1mm, 2mm, and 3mm, which will present peaks for frequencies of 0.5c/p, 0.25 c/p, and 0.16c/p respectively. Where 0.5c/p would represent high-frequency zones, that is, areas with very high detail, while 0.16c/p would present zones with lower or coarser frequencies.
If we apply this situation to an image or channel with a certain presence of noise, we will obtain a classification of it by order of frequencies and presence. In random noise, it will be distributed more or less uniformly across the entire set of frequencies, so the histogram will help us more to make decisions about its nature or incidence than the frequency spectrum.
On the other hand, the frequency domain can help us detect anomalies in a random noise distribution, anomalies whose origin may be noise reduction filters, image resampling, or other image manipulations that can give rise to certain patterns. Therefore, analysis through the frequency spectrum not only informs us about how various algorithms such as those associated with noise reduction or lossy compression are acting, but also about issues related to forensic image analysis, as it reveals inconsistencies in the construction of the image, such as patterns in areas that should a priori be random, different types of noise, or their presence in each channel.
Application of analysis in the frequency spectrum
Below are some examples of situations that we can highlight through image analysis via the frequency spectrum, which can greatly assist in forensic image analysis, quality control, or equipment characterization:
1st Presence of noise reduction filters
Usually, many processing tools introduce noise reduction filters in the image, for example, when strong white balances are made, balancing channels with low signal, the increase of noise in particular channels can be mitigated with a filter that produces smoothing of it. In this case, it is common to evidence the presence of low frequencies and absence of high frequencies.
2nd Presence of repeated or moved areas
When an area of the image is repeated, either by cloning or copying the same fragment multiple times, certain patterns are formed that completely break the randomness of the noise.
3rd Presence of lossy compression algorithms
A characteristic of many lossy compression algorithms is the ability to eliminate high frequencies, which are not relevant for the comprehension of the image by the human visual system. Thus, the spectrum will be rich in low frequencies and will decrease as we move toward high frequencies.
Practical case
A very common situation occurs when a strong adjustment is made to the color temperature of an image, for example, from a very warm illuminant at 3800K to a more neutral one around 5000K.
In this situation, the blue channel presents a very low signal level, so after the adjustment, it is necessary to apply a gain to this channel, which will reduce the signal-to-noise ratio (SNR) increasing the noise in that channel. Some tools automatically apply a noise reduction filter to that channel to mitigate the aesthetic effects of noise in areas associated with that channel.
These are very relevant situations for forensic image analysis since it is easy to evidence characteristic behaviors of tools, devices, and capture conditions, for example, deducing the illuminant of a scene even after white balancing.
Conclusions
Analysis in the frequency domain provides a perspective that correlates very well with certain structures in the image; therefore, it is a fundamental tool in the study of digital images, particularly in cases of forensic analysis or quality diagnostics.
This technique is advisable to perform in comparison with the spatial domain, that is, frequency spectrum versus histogram, since they are complementary tools; that is why in imageQA, both tools have been combined.
