# The Retinex algorithm for beautifying pictures

There are likely thousands of different algorithms out in the ether to “enhance” images. Many are just “improvements” of existing algorithms, and offer a “better” algorithm – better in the eyes of the beholder of course. Few are tested in any extensive manner, for that would require subjective, qualitative experiments. Retinex is a strange little algorithm, and like so many “enhancement” algorithms is often plagued by being described in a too “mathy” manner. The term Retinex was coined by Edwin Land [2] to describe the theoretical need for three independent colour channels to describe colour constancy. The word was a contraction or “retina”, and “cortex”. There is an exceptional article [3] on the colour theory written by McCann which can be found here.

The Retinex theory was introduced by Land and McCann [1] in 1971 and is based on the assumption of a Mondrian world, referring to the paintings by the dutch painter Piet Mondrian. Land and McCann argue that human color sensation appears to be independent of the amount of light, that is the measured intensity, coming from observed surfaces [1]. Therefore, Land and McCann suspect an underlying characteristic guiding human color sensation [1].

There are many differing algorithms for implementing Retinex. The algorithm illustrated here can be found in the image processing software `ImageJ`. This algorithm for Retinex is based on the multiscale retinex with colour restoration algorithm (MSRCR) – it combines colour constancy with local contrast enhancement. In reality it’s quite a complex little algorithm with four parameters, as shown in Figure 1.

• The Level specifies the distribution of the [Gaussian] blurring used in the algorithm.
• Uniform treats all image intensities similarly.
• Low enhances dark regions in the image.
• High enhances bright regions in the image.
• The Scale specifies the depth of the Retinex effect
• The minimum value is 16, a value providing gross, unrefined filtering. The maximum value is 250. Optimal and default value is 240.
• The Scale division specifies the number of iterations of the multiscale filter.
• The minimum required is 3. Choosing 1 or 2 removes the multiscale characteristic and the algorithm defaults to a single scale Retinex filtering. A value that is too high tends to introduce noise in the image.
• The Dynamic adjusts the colour of the result, with large valued producing less saturated images.
• Extremely image dependent, and may require tweaking.

The thing with Retinex, like so many of its enhancement brethren is that the quality of the resulting image is largely dependent on the person viewing it. Consider the following, fairly innocuous picture of some clover blooms in a grassy cliff, with rock outcroppings below (Figure 2). There is a level of one-ness about the picture, i.e. perceptual attention is drawn to the purple flowers, the grass is secondary, and the rock, tertiary. There is very little in the way of contrast in this image.

The algorithm is suppose to be able to do miraculous things, but that does involve a *lot* of tweaking the parameters. The best approach is actually to use the default parameters. Figure 3 shows Figure 2 processed with the default values shown in Figure 1. The image appears to have a lot more contrast in it, and in some cases features in the image have increased their acuity.

I don’t find these processed images are all that useful when used by themselves, however averaging the image with the original produces an image with a more subdued contrast (see Figure 4), having features with increased sharpness.

What about the Low and High versions? Examples are shown below in Figures 5 and 6, for the Low and High settings respectively (with the other parameters used as default). The Low setting produces an image full of contrast in the low intensity regions.

Retinex is quite a good algorithm for dealing with suppressing shadows in images, although even here there needs to be some serious post-processing in order to create an aesthetically pleasing. The picture in Figure 7 shows a severe shadow in a inner-city photograph of Bern (Switzerland). Using the Low setting, the shadow is suppressed (Figure 8), but the algorithm processes the whole image, so other details such as the sky are affected. That aside, it has restored the objects hidden in the shadow quite nicely.

In reality, Retinex acts like any other filter, and the results are only useful if they invoke some sense of aesthetic appeal. Getting the write aesthetic often involves quite a bit of parameter manipulation.

1. Land, E.H., McCann, J.J., ” Lightness and retinex theory”, Journal of the Optical Society of America, 61(1), pp. 1-11 (1971).
2. Land, E., “The Retinex,” American Scientist, 52, pp.247-264 (1964).
3. McCann, J.J., “Retinex at 50: color theory and spatial algorithms, a review“, Journal of Electronic Imaging, 26(3), 031204 (2017)

# Image sharpening – image content and filter types

Using a sharpening filter is really contingent upon the content of an image. Increasing the size of a filter may have some impact, but it may also have no perceptible impact – what-so-ever. Consider the following photograph of the front of a homewares store taken in Oslo.

The image (which is 1500×2000 pixels – down sampled from a 12MP image) contains a lot of fine details, from the stores signage, to small objects in the window, text throughout the image, and even the lines on the pavement. So sharpening would have an impact on the visual acuity of this image. Here is the image sharpened using the “Unsharp Mask” filter in ImageJ (radius=10, mask-weight=0.3). You can see the image has been sharpened, as much by the increase in contrast than anything else.

Here is a close-up of two regions, showing how increasing the sharpness has effectively increased the contrast.

Now consider an image of a landscape (also from a trip to Norway). Landscape photographs tend to lack the same type of detail found in urban photographs, so sharpening will have a different effect on these types of image. The impact of sharpening will be reduced in most of the image, and will really only manifest itself in the very thin linear structures, such as the trees.

Sharpening tends to work best on features of interest with existing contrast between the feature and its surrounding area. Features that are too thin can sometimes become distorted. Indeed sometimes large photographs do not need any sharpening, because the human eye has the ability to interpret the details in the photograph, and increasing sharpness may just distort that. Again this is one of the reasons image processing relies heavily on aesthetic appeal. Here is the image sharpened using the same parameters as the previous example:

There is a small change in contrast, most noticeable in the linear structures, such as the birch trees.  Again the filter uses contrast to improve acuity (Note that if the filter were small, say with a radius of 3 pixels, the result would be minimal). Here is a close-up of two regions.

Note that the type of filter also impacts the quality of the sharpening. Compare the above results with those of the ImageJ “Sharpen” filter, which uses a kernel of the form:

Notice that the “Sharpen” filter produces more detail, but at the expense of possibly overshooting some regions in the image, and making the image appear grainy. There is such as thing as too much sharpening.

So in conclusion, the aesthetic appeal of an image which has been sharpened is a combination of the type of filter used, the strength/size of the filter, and the content of the image.

# Why image processing is an art

There are lots of blogs that extol some piece of code that does some type of “image processing”. Classically this is some type of image enhancement – an attempt to improve the aesthetics of an image. But the problem with image processing is that there are aspects of if that are not really a science. Image processing is an art fundamentally because the quality of the outcome is often intrinsically linked to an individuals visual preferences. Some will say the operations used in image processing are inherently scientific because they are derived using mathematical formula. But so are paint colours. Paint is made from chemical substances, and deriving a particular colour is nothing more than a mathematical formula for combining different paint colours. We’re really talking about processing here, and not analysis (operations like segmentation). So what forms of processing are artistic?

1. Anything that is termed a “filter”. The Instagram-type filters that make an ordinary photo look like a Polaroid.
2. Anything with the word enhancement in it. This is an extremely loose term – for it literally means “an increase in quality” – what does this mean to different people? This could involve improving the contrast in an image, removing blur through sharpening, or maybe suppressing noise artifacts.

These processes are partially artistic because there is no tried-and-true method of determining whether the processing has resulted in an improvement in the quality of the image. Take an image, improve its contrast. Does it have a greater aesthetic appeal? Are the colours more vibrant? Do vibrant colours contribute to aesthetic appeal? Are the blues really blue?

Consider the photograph above. To some, the image on the left suffers from being somewhat underexposed, i.e. dark. The image in the middle is the same image processed using a filter called Retinex. Retinex helps remove unfavourable illumination conditions – the result is not perfect, however the filter can help recover detail from an image in which it is enveloped in darkness. Whilst a good portion of the image has been “lightened”, the overcast sky has darkened through the process. There is no exact science for “automagically” making an image have greater aesthetic appeal. The art of image processing often requires tweaking settings, and adjusting the image until it appears to have improved visually. In the final image of the sequence below, the original and Retinex processed images are used to create a composite by retaining only the maximum value at each pixel location. The result is a brighter, contrasty, more visually appealing image.

# Image enhancement (4) : Contrast enhancement

Contrast enhancement is applied to images where there is a lack of “contrast”. Lack of contrast manifests itself as a dull or lacklustre appearance, and can often be identified in image histograms.  Improving contrast, and making an image more visually (or aesthetically) appealing is incredibly challenging. This is in part because the result of contrast enhancement truly is a very subjective thing. This is even more relevant with colour images, as modifications to a colour, can impact different people differently. What ideal colour green should trees be? Here is a brief example grayscale image and its intensity histogram.

It is clear from the histogram that the intensity values do not span the entire range of values, effectively reducing the contrast in the image. Some parts of the image that could be brighter, are dull, and other parts of the image that could be darker, are lightened. Stretching both ends of the histogram out, effectively improves the contrast in the image.

This is the simplest way of enhancing the contrast of an image, although the level of contrast enhancement applied is always guided by the visual perception of the person performing the enhancement.

# Image enhancement (3) : Noise suppression

Noise suppression may be one of the most relevant realms of image enhancement. There are all kinds of noise, and even digital photographs are not immune to it. Usually the algorithms that deal with noise are grouped into two categories: those that deal with spurious noise (often called shot or impulse noise), and those that deal with noise that can envelop a whole image (in the guise of Gaussian-type noise). A good example of the latter is the “film grain” often found in old photographs. Some might think this is not “true” noise, but it does detract from the visual quality of the image, so should be considered as such. In reality noise suppression is not as important in enhancing images from digital cameras because a lot of effort has been placed on in-camera noise suppression.

Below is an example of an image with Gaussian noise. This type of noise can be challenging to suppress because it is “ingrained” in the structure of the image.

Here are some different attempts at trying  to suppress the noise in the image using different algorithms (many of these algorithms can be found as plug-ins to the software ImageJ):

• A Gaussian blurring filter (σ=3)
• The Perona-Malik Anisotropic Diffusion filter
• Selective mean filter

To show the results, we will look at the extracted regions from some of the algorithmic results compared to the original noisy image:

It is clear the best results are from the Perona-Malik Anisotropic Diffusion filter [1], which has suppressed the noise whilst preserving the outlines of the major objects in the image. The median filter has performed second best, although there is some blurring which has occurred in the processed image, which letters in the poster starting to merge together. Lastly, the Gaussian blurring has obviously suppressed the noise, whilst incorporating significant blur into the image.

Suppressing noise in an image is not a trivial task. Sometimes it is a tradeoff between the severity of the noise, and the potential to blur out fine details.

[1] Perona, P.,  Malik, J., “Scale-space and edge detection using anisotropic diffusion”, In: Proceedings of IEEE Computer Society Workshop on Computer Vision,. pp.16–22. (1987)

# Image enhancement (2) : the fine details (i.e. sharpening)

More important than most things in photography is acuity – which is really just a fancy word for sharpness, or even image crispness. Photographs can be blurry for a number of reasons, but usually they are all trumped by lack of proper focusing, which adds a softness to an image. Now in a 3000×4000 pixel image, this blurriness may not be that apparent – and will only manifest itself when an enlargement is made of a section of the image. In terms of photographing landscapes, the overall details in the image may be crisp, however small objects may “seem” blurry, because they are small, and lack detail in any case. Sharpening will also fail to fix large blur artifacts – i.e. it’s not going to remove defocus from a photograph which was not properly focused. It is ideal for making fine details crisper.

Photo apps and “image editing” software often contains some means of improving the sharpness of images. Usually by means of the “cheapest” algorithm in existence – “unsharp masking”. It works by subtracting  a “softened” copy of an image from the original. And by softened, I mean blurred. It basically reduces the lower frequency components of the image. But it is no magical panacea. If there is noise in an image, it too will be attenuated. The benefit of sharpening can often be seen best on images containing fine details. Here are examples of three different types of sharpening algorithms on an image with a lot of fine detail.

Three filters are shown here are (i) Unsharp masking (USM), (ii) Cubic Unsharp masking (CUSM) and (iii) Morphological sharpening (MS). Each of these techniques has its benefits and drawbacks, and the final image with improved acuity can only really be judged through visual assessment. Some algorithms may be more attune to sharpening large nonuniform regions (MS), whilst others (USM, CUSM) may be more aligned with sharpening fine details.

# Image enhancement (1) : The basics

Image enhancement involves improving the perceived quality of an image, either for the purpose of aesthetic appeal, or for further processing. Therefore you are either enhancing features within an image, or suppressing artifacts. The basic forms of enhancement include:

• contrast enhancement: enhancing the overall contrast of an image, to improve dynamic range of intensities.
• noise suppression: reducing the effect of noise contained within an image
• sharpening: improving the acuity of features within an image.

These relate to both monochromatic grayscale and colour images (there are additional mechanisms for colour images to deal with enhancing colour). The trick with these enhancement mechanisms is determining when they have achieved the required effect. In image processing this is often a case of the rest being “in the eye of the beholder”. A photograph who’s colour has been enriched may seem pleasing to one person, and over saturated to another. To illustrate, consider the following example. This image is an 8-bit image that is 539×699 pixels in size.

Here is the histogram of its pixel intensities:

From both the image and histogram, it is possible to discern that the image lacks contrast, with the majority of gray intensities situated between values 25 and 195. So one of the enhancements could be to improve its contrast. Here is the result of a simple histogram stretching:

It may then be interesting to smooth noise in the image or, sharpen the image to enhance the letters in the advertising. The sub-image extracted from the above shows three different techniques.

# The perception of enhanced colour images

Image processing becomes more difficult when you involve colour images. That’s primarily because there is more data involved. With monochrome images, there is really only intensity. With colour images comes chromaticity – and the possibility of modifying the intrinsic colours within an image whilst performing some form of image enhancement. Often, image enhancement in colour images is challenging because the impact of the enhancement is very subjective.

Consider this image of Schynige Platte in Switzerland. It is very colourful, and seems quite vibrant.

The sky however seems too aquamarine. The whole picture seems like some sort of “antique photo filter” has been applied to it. How do we enhance it, and what do we want to enhance? Do we want to make the colours more vibrant? Do we want to improve the contrast?

In the first instance, we merely stretch the histogram to reduce the gray tonality of the image. Everything becomes much brighter, and there is a slight improvement in contrast. There are parts of the image that do seem too yellow, but it is hard to know whether this is an artifact of the original scene, or the photograph (likely an artifact of dispersing yellow flower petals).

Alternatively, we can improve the images contrast. In this case, this is achieved by applying a Retinex filter to the image, and then taking the average of the filter result and the original image. The resulting image is not as “bright”, but shows more contrast, especially in the meadows.

Are either of these enhanced images better? The answer of course is in the eye of the beholder. All three images have certain qualities which are appealing. At the end of the day, improving the aesthetic appeal of a colour image is not an easy task, and there is no “best” algorithm.