The Head and Bill comp is proving to be quite a challenge for most competitors. The main issue is getting close enough to the bird to be able to make a tight enough crop to meet the competition theme. Tight cropping has its own issues, notably, digital noise. The more one zooms in the more the digital noise becomes a problem. An advanced noise reduction algorithm can make a big difference; those based on or derivative of Wavelet Theory work best. The one I prefer is Neat Image as it satisfies one of the basic requirements of effective noise reduction; it does not destroy image detail even when strong NR is applied.

The image of a back-illuminated razor blade is an ideal test object with which to study noise reduction and sharpening halos. Also, it can be used to measure the camera + lens Spatial Frequency Response (SFR) and lens Modulation Transfer Function (MTF). In digital images there are two types of noise; photon shot noise and electronic noise. Photon shot noise is related to the particle nature of photons and their random rate of arrival at a particular sensor pixel. Over the entire array of sensor pixels there will be a statistical fluctuation in the number of photons detected during the exposure time. The result of this is that the image will have a speckle-like appearance that we see in many images, especially in soft out of focus backgrounds and sky. The contrast variation is random and its histogram has a mathematical form called a Poisson Distribution. This photon shot noise is obvious in the bright parts of most images when we zoom in for a closer look or make a big crop. It is the most common type of digital noise in our images and any noise reduction software worth having must do a good job at reducing this kind of noise. Electronic noise has its origin in the photodiodes of the sensor pixels and associated amplifiers, analogue to digital converters, and data read-out circuits. The electronic noise is most obvious in the dark parts of the image where the contribution from photons in minimal and electronic noise is the major noise source. If we amplify the brightness recorded in the dark parts of an image by, for example, raising the shadow detail, then we will start to see the electronic noise. The noise statistics of the electronic noise are not usually the same as photon shot noise so the noise reduction applied to the dark parts of the image will be different to that applied to the bright parts. Good noise reduction software is able to detect the different noise profiles in the light and dark parts and apply the optimal amount of noise reduction to those parts. Below I show an example; the right-hand image of a razor blade edge is the baseline image with some capture sharpening in DPP4 but no noise reduction. In the bright part of the image the photon shot noise is easy to see and I have measured the luminosity signal to noise ratio (SNR). The amount of noise is not as bad as it looks because this is a 400% crop. On the dark side the SNR=675 which is very good. It is generally agreed that acceptable image quality needs SNR>30. The middle image shows the same crop but with the default noise reduction amount determined by Neat Image and no sharpening has been applied by Neat Image. The left-hand image shows the situation after fairly strong luminance noise reduction has been applied and no sharpening by Neat Image. An important observation is that the halo contrast on the hard edge appears about the same for each of these images suggesting the sharpness has been preserved even in the left-hand image where strong NR has been applied. More on this below.

I noted noise reduction software should not only do a good job on the noise but it should preserve the contrast of fine detail, in other words, the applied noise reduction should not affect the sharpness of the image. The sharpness of an image is best characterized using the system SFR. The SFR is the contrast in the digital image measured over a range of spatial frequencies (cycles/mm) from zero up to the sampling frequency. Fine detail in an image contains a lot of higher spatial frequencies and it is these that we must preserve during noise reduction. An important feature of the SFR is that when SFR=1, this means the contrast in the digital image equals the contrast of the object we are photographing. This is the ideal situation but our camera systems are far from perfect so this condition cannot be achieved over the entire range of spatial frequencies found in the object. However, the SFR curve does show us when SFR>1 indicating over-sharpening which is handy to know when evaluating the performance of software providing noise reduction and sharpening options. These points are illustrated in the SFR graphs below which were measured using the three razor blade images already discussed. The main point to note is that the three graphs are very similar and any differences are of no practical significance. The fact that the graphs are so similar indicates that the application of the Neat Image noise reduction has not affected the sharpness and contrast of fine detail, a basic outcome of effective noise reduction.