Welcome back. When the noise in an image is additive, but not broadband, but instead it's impulsive, also referred to as data dropout or, and spike noise, then the LSA filters we covered in the previous segment are not appropriate. If applied, the resulting image looks worse, since the impulsive noise is spread out. A specific form of impulsive noise is salt and pepper noise, called so since it resembles granules of salt and pepper randomly spread over the image. For impulsive noise, a class of non-linear filters, referred to as order statistic filters, are appropriate. A specific example of such filter is the median filter. We discussed median filters, one-dimensional, two-dimensional, multi-dimensional, and general, in this segment and show a number of results of its application to images degraded by impulsive noise. We'll also discuss other order statistic filters, such as the alpha-trimmed mean filter and non-linear filters with mean in their name, but not the arithmetic mean filter, which is an LSA filter, and we covered this in the last segment, but for example the geometric mean or the harmonic mean filter. We also show some experimental results when additive broadband and additive impulsive noises are combined. So let us proceed with this exciting material. A commonly encountered filter in image enhancement is the median filter Let's see how it is defined. Let x be the input image, y the output image. So, the value of y at location n1, n2 can be found by considering a neighborhood around the pixel n1, n2 of the input image. Then, take the values inside this neighborhood and order them in increasing or decreasing fashion. And then pick the middle value if the number of pixels is odd or the average of the two middle values, if the number of pixels is even. So it's a rather straightforward operation in, in figuring out the, the median filter. Let's look at some of the properties. The filter is non-linear. And this is straightforward to confirm by utilizing the definition of linearity that we gave earlier. Because of that, clearly I cannot describe this filter in the frequency domain, as was the case with the min filter we studied earlier. This filter is very useful for reducing impulsive, or also referred to as salt and pepper noise. By the way, this is also referred to as data drop out or spike noise. So it's generated if I take a value of the input image considering all the values so that the value of n1, n2. If it changes to a value a with probability P of a and the value at n1, n2 changes to value b with probability P of b, then if both P of a, P of b are non-zero, we talk about the bipolar type of noise, and salt and pepper is another name for that. Typically when a equals to 255 and b equals to zero, and clearly resembles throwing salt and pepper granules randomly over the image, if one of the P of a, P of b is equal to zero, then it's a unipolar noise, and then it can be referred to as salt or pepper, just salt or just pepper noise. So the median filter, as we'll show right way with a toy example, is again suitable in removing this type of noise, while preserving the spacial resolution, not introducing blurring, and also preserving the edges. The filter however is not very good if the noise is broadband, as were the case in what we studied earlier, with respect to the min filter. Finally, as we'll show with a toy example there, the one-dimensional median has a better performance in preserving edges in an image than the two-dimensional median. And very often we'll, we'll implement the two-dimensional median as a separable operation to one-dimensional median in the vertical and the horizontal directions. Let us look at the application of a three-point one-dimensional median filter, filter to a simple signal. So first of all to find the location of the output at n we'll considering input values at location n, n-1 and n+1. So for this simple signal here, by looking at it we can say that this particular value is noisy. It does not fit well with the neighboring values. So to find the output of the median filter at this location, I consider the values of the input, 1, 2, 1. Their median is 1. So therefore, the noise is eliminated. If I keep processing, also I see easily that this perfect edge here of the input image's signal is finally preserved. So here is my y of n. Noise is eliminated, sharp edges preserved, and this is the main characteristic of the median filter. We show here a simple example of the application of a two-dimensional, now, three by three median filter. So here is the image I want to process. So the solid dots have a value of one, and the thin dot a value of zero. So this is a perfect three by three square sitting on the zero level around it. So if I consider the processing of this pixel of the input, then I have to find the median of these values. And the median, the value in the middle, is equal to zero. So we see that that edge pixel there of the input image has been eliminated, and this square image of the input will end up being this cross-shaped picture in the output. So the corners are distorted, are not preserved, and therefore, we typically prefer to use separable median filters. So first filter the image in the vertical direction with a one-d median, and then the resulting image is filtered in the horizontal direction by another one-dimensional median. We see here an example of median filtering. Here is the original image. Here is the noisy one, corrupted by salt and pepper noise. This is the result of the application of a three by three median, and this is the result of the application of a five by five median. In both cases, the noise has been completely eliminated, however, with a five by five median the image has been considerably blurred, and therefore, for this particular experiment the choice, our choice, should be the three by three median. Let us look now at some additional examples of the application of median filtering. We show here a colored image. It's corrupted by salt and pepper noise, and this noise has been added independently to the three color channels, the R, G, and B channels. 10% of the pixels are corrupted by salt and pepper noise. If we first use a noise smoothing filter that has Gaussian shape, a five by five filter, then we obtain this result. Clearly the impulses, the impulsive noise has been spread out by this noise smoothing filter. And this does not seem the appropriate filter to be utilized in this particular case. If we now use a median filter, a three by three median filter, then this is the result we obtain. Clearly, most, if not all, of the spikes of the impulses have been removed, while the sharpness of the image has been maintained. So we compare the previous result with this one, it's certainly a major improvement. In this example, we utilize the same image again, but the noise is heavier now, 20% of salt and pepper noise. So if we apply the Gaussian filter, this is the result we obtain. Again, the impulses are spread around, therefore this is not the appropriate filter to use. I should mention here that, on one hand ,the noise is added on each color channel separately, and the filter's applied on each color channel separately. So if I apply now a three by three median filter, this is the result we obtain. The noise is almost removed, however there are still impulses that can be seen in the processed image. So in this particular case we apply the median filter a second time, and this is the result, further improved, although there are some impulses around, for example here. And we can apply it one more time, so after the third pass this is the result. The impulsive noise is almost completely removed. However, in each application of the median filter, there is some blurring that occurs, that's to, to the image. So the, the motivation for this example is just to show the simple fact that the median filter can be applied multiple times. If the noise is heavy and is not eliminated to a satisfactory degree by the first pass, we can keep repeating it as it was demonstrated through this example. In addition to the median filter, which you studied in some detail, there are a number of other filters under the grouping of order statistic filters suitable primarily for handling impulsive type of noise. They're called so because they're based on the ordering of the pixel values in the neighborhood. So here's again the expression for the median filter, written slightly differently. So x is the input image, y the output. So the value at n1, n2 is the result, is the median, of the values of the input in a neighborhood N which is centered at n1 and n2. Another filter in this family is the max filter, which finds the brightest, the maximum value, the brightest intensity in the neighborhood. Since pepper has low value it's reduced by this operation. It's replaced by the maximum value in the neighborhood, so this is therefore good for handling pepper type of noise. [BLANK_AUDIO] The min filter finds the darkest point in the neighborhood, the minimum value, and therefore, this is good for reducing salt noise. [BLANK_AUDIO] The mid-point combines order statistics and averaging. So it is the average of the maximum and the minimum values in the neighborhood. And this works best for randomly distributed noise like Gaussian or uniform. And finally, the alpha-trimmed mean. So from the, all the pixels in the neighborhood, we delete the alpha over two lowest and alpha over two highest values. And then for the remaining values, which number the number of pixels in the neighborhood minus alpha, which I have deleted, we find the mean. So clearly for alpha equals zero we obtain the arithmetic mean filter. [BLANK_AUDIO] While for alpha equals the number of pixels in the neighborhood minus one, we obtain the median filter. [BLANK_AUDIO] For the other values of alpha, we obtain filters which are useful for situations that involve multiple types of noise, such as a combination of salt and pepper and Gaussian noise. There's a family of filters that have the term, mean, in their name. The first one is arithmetic mean filter, which you've studied extensively. We've also been referring to it by the name of average or flat filter. x, again, is the input to the filter, and y is the output. So to find the output value at n1, n2, I'm considering the pixels in the neighborhood N which is centered at n1, n2. And then I sum the pixels up in this neighborhood and divide by that number. I find the arithmetic mean. So this here denotes the number of pixels in the neighborhood, or the cardinality of the set of pixels in the neighborhood N. The geometric mean filter finds the geometric mean of a set of numbers, so it takes the samples in the neighborhood and multiplies them, and then this is the nth root, where again, this is the Number of pixels in the neighborhood. The geometric mean filter has performance comparable to the mean, but by and large tends to lose less detail in the process. The harmonic filter is shown here. This does well with Gaussian type of noise and also for salt type of noise but not for pepper type of noise. [BLANK_AUDIO] The final filter is the contra-harmonic mean. Q is the order of the filter. [BLANK_AUDIO] This filter for Q greater than zero eliminates [BLANK_AUDIO] pepper type of noise. And for Q less than zero it eliminates salt, but not clearly able to eliminate both at the same time. It is verified that for Q equals zero, I obtain the arithmetic mean. [BLANK_AUDIO] While for Q equals minus one, I obtain the harmonic mean filter. [BLANK_AUDIO] With the exception of arithmetic mean filter, the remaining are non-linear filters, and therefore, I cannot really describe them in the frequency domain, see their shape in the frequency domain. Let's look at an example now where the image has been corrupted by two types of noise. So this circuit board image is corrupted by additive uniform noise, as shown here. And then on top of that, salt and pepper noise is added with the result shown here. So it's a 10% salt and pepper noise. So we want to apply some of the filters we just discussed to the image again that has a combination of these two noises. And if we use an arithmetic mean filter of five by five, we obtain this result, while a five by five geometric provides this result. Both results are really not that good, not maybe acceptable, and certainly the one by the geometric mean is not good at all. Now if we apply a five by five median we obtain this image, while a five by five alpha-trimmed mean with alpha equals five provides this result. So in this case, both results are considerably better, and quite good. And one might even say, by looking at the fine details, that the alpha-trimmed mean is slightly better than the median filter. Again, this is being an enhancement problem, there are no guidelines says which filter and with which parameters is going to provide the better, the best results. Instead we try the various filter that we have in our ammunition. We also change their parameters appropriately. If there is a lot of noise then the support, for example, becomes larger until we find the result that is acceptable, either for human appreciation or as an input to a followup processing step. [BLANK_AUDIO]
