Neurocomputing Device for Image Processing and Classification

This article is concerned with problems of construction of specialized computing device for image processing and classification. The device may be used for classification of objects images in conditions of various image scales of the same object. The device may be realized by neurochips technology.


INTRODUCTION
The problem of classification of the object images be the area of the k th object, knowing that objects images don't intersect.
A set of background images S (x,y) and separate different objects 0 ( , ), 1, k S x y k m = forms the image: (3) Satisfying the conditions 1-3 means that: On solving the problem of classification performed on the image S(x,y) of multi m 0 objects, into unknown in advance number of various classes. First stage of processing is procedure of segmentation of the image S(x,y), mathematically segmentation procedure may be described by the evaluation of the following predicate: The result of predicate (4), the source image in equation (3)  = ; we get the following: We see that the ratio (5) is invariant to the change of square scale. Now let us show that the invariant of the ration (5) to the scale changes for any shape. In Fig.  1 is shown that any figure may be shown as a sum of squares the area of which aims to zero. For shape A(x,y), the ratio (5) may be given in the following equation: where 1 , q a a -are the length of the first and last squares; ( 1, ) -are the width of squares. On changing the scale of the image A(x,y) in k times, we have Thus, the ratio (5) is invariant for any shape presented in different scale.
To calculate the area of k-object image, which is in the area of vision D(x,y) using predicate (4) it is enough to calculate the number N k of pixels forming the image and multiply it by the area S p of one pixel: .
To calculate the perimeter of any k image, we consider that it consists of the length of separate pixels. As shown in Fig. 2 any pixel of the image may have one of the following numbers of adjacent pixels.
Pixels marked on Fig. 2 by the digit 1, have 4 neighboring pixels (up, down, right, left), the contribution of theses pixels to the length of perimeter is equal zero.
Pixels marked on Fig. 2 by the digit 2, each have three neighboring pixels; their contribution to the perimeter length is equal to the length of one pixel 1d.
Pixels marked on Fig. 2 by the digit 3, they have two neighboring pixels; their contribution to the perimeter length is equal to 2d.
Pixels marked on Fig. 2 by the digit 4 have only one neighboring pixel; their contribution to the perimeter length is equal to 3d.
Pixels marked on Fig. 2 by the digit 5; their contribution to the perimeter length is equal to 4d.
By analyzing Fig. 2 and the given reasoning, we can calculate the length of perimeter of any image by the following: Where , 1,3 r N r = is the number of excited neurons which have r neighboring excited pixels in the image; N 0 is the number of neurons which don't have neighboring pixels in their space.
Images classification using ratio (5), it is rational to be carried out with the help of technology of neural networks. One of the possible architecture is given in Fig. 3. The device has the area of vision D(x,y) from matrix n 1 ×n 2 of binary x neurons, where n 1 -number of rows, n 2 -number of columns. Every x-neuron corresponds to one pixel of the image D(x,y) .
The function of X-neurons may be described by the following formula: Where U in ; U out -are correspondingly input and output signals of X-neurons; n θ -Threshold value of xneurons.   ( 1 1), ( 1 ), ( 1 1), The value of output signals from neurons of field Z k may be determined by the following formula: where k 1 -is a positive constant; Thus, the output signal will be signal proportional to the general number of the excited x-neurons and the area of given image. When output signal of a neuron ( 1, 4) q q = is proportional to the number of elements of the image which have from q to four excited conditions. Using the equation (9), it is easy to receive the signals proportional to number N 0 of neurons which don't have neighboring elements on the image and signals proportional to numbers N r ( 1,3) r = of neurons which has correspondingly from one to three excited elements in the image Having signals proportional to the area and perimeter of the image, it is easy to receive signal as well with the help of ALU block, which will be proportional to the first part of equation (5).the resultant value of signal may be used to classify input image.

RESULTS
Experiments conducted on image like given in Fig.  4 prove the ability to work with the offered algorithm. The developed device is suggested to apply for speeding up image process in different fields of applications.