Poster
# 62
Poster |
|
6th Internet World Congress for Biomedical Sciences |
Gustavo Camps i Valls(1), Antonio J. Serrano López(2), Jesús Modia(3), José David Martín Guerrero(4), José Vicente Francés Víllora(5)
(1)Universitat de València - Burjassot, Valencia. Spain
(2)(3)Dpto. Electrónica. Universidad de Valencia - Burjasot. Spain
(4)G.P.D.S. Departament d´Enginyeria Electrònica. Universitat de València - Burjassot. Spain
(5)Dpto. Ingenieria Electrónica. University of Valencia - Burjassot. Spain
|
 |
|
|
[Health Informatics] |
[Medical Electronics & Engineering] |
Any image acquired by a device is susceptible of being degraded by the environment of acquisition; phenomenons of noise of the sensor can appear, blurring because of an unfocused objective, movement, variations in the illumination, etc. The restoration of images tries to minimize the effects of these degradations by means of a filtrate. It is distinguished from enhancement techniques in that those are in charge of more for the accentuation and extraction of characteristics than of their restoration.
Therefore, a fundamental problem in the image processing is the improvement of their quality through the reduction of the noise that they can contain being often known as "cleaning of images". A great variety of techniques dedicated to carry out this task exists and each one of them is focused on different features from the image and the noise.
When the noise has some characteristic pattern not completely aleatory in time, and therefore it doesn´t present an analyzable behavior from the statistical point of view, it is much more difficult to establish numeric describers on the quantity of noise. However, they are used the same techniques approximately in the elimination of the noise producing a smaller effectiveness.
In the field of images restoration it is assumed, in most of methods, that pixels in the image are much smaller that any detail and that many of their neighbors represent the same structure. With these assumptions, it is possible to apply methods based on substituting any pixel with its averaged vicinity to restore or to clean images with random noise.
A great variety of techniques dedicated to carry out this task exists and each one of them is guided to different characteristics from the image and the noise. With some different suppositions and a completely particular way is approached each problem: this will make an useful technique in the elimination of certain characteristics of noise not to be it so much before others. The simplest procedure of averaging spacely is simply made by adding the values of brightness of each píxel in each region of the image and to divide by the number of pixels of the environment, using this result in the construction of a new image.
The vicinity operations that include the multiplication for kernels are usually applied using symmetry around the pixel. This creates a problem for the pixels near the borders of the image since they have half of neighboring pixels. To solve this sort of problems many techniques have been devised: starting from masks or asymmetric kernels to special marginal conditions.
Varying the values of the weights of these kernels then a set of masks more or less effective in the cleaning of the noise will be found but they will have as a common effect, and at the same time undesired, the blurring of the image; this is the case of the well-known median filter.
The main function of the median filters is to force the points with values of gray very different to its neighbors to have next values to them, so that picks of intensity that appear in uniform areas are eliminated. The median filters have been broadly used in the elimination of the impulsional noise (noises of maximum or minimum value). Its ability to eliminate impulses grows with the size of the window of the filter but its main drawback is to blur the images, losing details and definition. In our work we will try to eliminate random noise on images with 256 levels of gray using a selective mask of medium.
Evidently the computacional burden increases with the width of the vicinity area. The algorithm that we present is based on a double outline: first it tries to identify and to locate the random noise in the image and in second place, the filter is applied of medium only under certain circumstances. Therefore it is a Selective Filter of Medium. The central idea of the procedure will be the one of marking all the points that belong to a chain of at least longitude equal to 3 (probably being part of the original image and not of the noise). Then the median filter will be applied on the points that are not marked that will correspond to the isolated plots, that is to say, to the noise. These two processes, the one of marking and the one of applying the medium one are made in fact simultaneously.
Differences in values of gray smaller than 18 demonstrated in general not to be detectables for the human eye and for that reason, the noise that doesn´t differ of the original image in bigger values that 18 won´t be necessary to eliminate it.
Evidently there is a relationship between the intensity of the noise and the ownership or not of these chains to the original image. When an intense noise invades an image it will cause the appearance of chains of noise of those longitudes. However it is not very probable that in an image with little noise chains of noise of longitudes 3 and 4 appear, for what you/they will very probably belong to the original image in this case and they should not be substituted therefore.
A way to identify the uniform areas is by means of a variation of the previous algorithm: the points that belong to uniform areas will belong to a great chain and therefore they will have a high value in this auxiliary matrix.The chains that we look for to eliminate will have values of 3 and of 4 and they will be surrounded by high values. This algorithm has been experienced with quite good results, although it is still in phase of improvement.
To measure the error made in the reconstruction of the images, i.e. the differences among the original and reconstructed image, different functions and norms have been taken: the quadratic error (ECM), signal-to-noise ratio (SNR) and the laplacian metric (HIM). But far away from the mathematical measures of similarity or error that provide us these formulas, we can also have subjective approaches or of optic appreciation of the image. It is in this aspect where our algorithm has a better performance, since it eliminates the noise accurately
To measure the error made in the reconstruction of the images, i.e. the differences among the original and reconstructed image, different functions and norms have been taken: the quadratic error (ECM), signal-to-noise ratio (SNR) and the laplacian metric (HIM). But far away from the mathematical measures of similarity or error that provide us these formulas, we can also have subjective approaches or of optic appreciation of the image. It is in this aspect where our algorithm has a better performance, since it eliminates the noise accurately
We can conclude that for a good cleaning and restoration of images it is evident that most of the work should be previous in the localization of the noise. Once this is effective for the image type and noise that we are working with, we will be able to apply the many existing tools of elimination of noise we want to, lineal or adaptative ones.
The proposed algorithm has been tested in a great number of clinical images revealling an excellent behaviour although some kind of empirical adjustments have to be made.
|
|
|
|
|
|
[Health Informatics] |
[Medical Electronics & Engineering] |
