July 02 Reviewed: August 2022

DISCUSSION
The scheme of image distortion and filtering is considered the simplest linear filtering algorithm based on local 
processing of a noisy image of small dimensions. A technology for filtering digital images by the convolution method 
with a pulse characteristic in the spectral region is proposed. Filtering of images distorted by Gaussian noise is carried 
out. A method of adaptive anisotropic filtering has been developed to eliminate noise in images. When implementing 
the described method, a set of pulse characteristics of all assigned forms of the length specified in the dialog is 
generated before filtering begins. It is shown that, in the absence of noise, the filtering result coincides with the initial 
brightness value of the filtered image. If the noise variance is large, the multiplier in front of the square brackets in 
the formula becomes very small, the brightness estimate approaches the average brightness value inside the window. 
The result of filtering is obtained better than with non-adaptive filtering. A modified adaptive method of noise 
elimination by a piecewise smooth image model is proposed. 
CONCLUSION
Fractal compression methods based on the elimination of structural redundancy can provide the required 
compression of the video stream by 150-200 times, but they have very low performance and currently cannot provide 
real-time processing of the video stream. 
Thus, to date, existing image processing algorithms can only achieve 130-150 times compression of the video 
stream due to a noticeable deterioration in their visual quality. Therefore, to ensure good quality of TV images with a 
frame size of 8-10 kBytes, it is necessary to develop new effective methods for processing video streams that 
significantly minimize the amount of metadata (no more than 500 bytes per frame), or do not use motion compensation 
at all. 
The sigmoidal activation function is most often used, since it is differentiable, has the property of amplifying weak 
signals and preventing saturation from large signals, since they correspond to the regions of arguments where the 
sigmoid has a gentle slope. 
From the point of view of architecture, the following main types of neural networks are distinguished: fully 
connected, multi-layered, with local connections. 
From the point of view of the neural network approach, the problem of image filtering in order to highlight the 
boundaries of objects can be considered as a problem of approximating the function of variables for a window size . 
It is known that a multilayer perceptron is able to form an arbitrary multidimensional function at the output (the 
main difficulty lies in choosing the number of layers and neurons), which characterizes it as a universal tool for 
approximating the function, and determines the possibility of using this type of network for the task of filtering noise 
in the image. In most of the noise reduction (filtering) methods used, the output image is formed by pixel. [13, 14].