Assesment techniques of the products on the basis of shape analysis.

Buzera, Marius ; Prostean, Gabriela ; Stefan, Constantin 等

1. INTRODUCTION

Shape is an essential parameter when analyzing the vegetal products, especially due to the fact that, by taking It into consideration on can appreciate the integrity of the products, their quality as well as their size. If in the case of industrial products analysis, it is but easy to make an assessment for vegetal products, having an endless number of possible shapes one can find it very hard to assess. (Baoping, 1999) Also the usage of classical algorithms under classification, irregardless of the calculus ability to use, is to became inefficient. (Tao & Wen, 1999)

Thus, the algorithms assessing the products throughout this article based on conexionist patterns--neural networks and they ensure the assessment of the quality of products according to the basic shape of the products under three classes: class A good products, class B--average products and class C--under standard products, according to the integrity of the products:

class A--good products, class B--damaged products and according to the size: class A--big products, class B--medium products and class C--small products.

The assessment of the quality of the products in question was carried out on an experimental stand, on the basis of machine vision principles.

2. THE EXPERIMENTAL DEVICE THE "MACHINE VISION" SYSTEM

The device of video-inspection has been carried-out around transporters with a belt activated by an electric engine, which ensures the moving of the product in face of the video-inspection system, made of two video cameras with manual focusing. The cameras are disposed in the same frame, realizing an angle of [60.sub.0] , towards the analyzed object (Laykin et al., 2003). These allow the acquisition of 4 coloured RGB pictures of [512.sup.*] 512 pixels, which they transmit in a real time, to the analyzing mechanism represented by a PC of Pentium IV/3200 MHz type.

[FIGURE 1 OMITTED]

For a better illumination of the product, the interior walls of the illuminating room were painted white, such that they can ensure a diffuse illumination of the object under inspection. (Laykin et al., 2002) Also, the illuminating room was implemented on a mobile metallic framework that allows to change the distance between electric bulbs and the object, and to reposition the electric bulbs on the prop, as shown in Fig. 1.

(Buzera et al, 2008)

3. THE SHAPE. THE ANALYSIS OF THE SHAPE

After the outline of the object had been carried-out, we must obtain the signature of the shape of the object. The signature of the shape is a feature that can be used to identify the objects. A main part in getting a right answer is that of the precision of the weigh centre. (Sudharkara, 2003)

For each image the shape of the object was automatically determined through an analytical way, due to the help of an application carried-out in the Matlab program. The polygonal line that defines the outline analytically and successively over crossed with a right line. The angle of orientation was considered of 50 (Buzera & Stefan, 2007).

In order to detect the shape of the object from the picture the most important information was represented by the position of the pixels that describe the outline. Other information can be ignored.

The signature of the shape of the object supposes that the weight-centre of the outline should be determined. These can be determined due to the moments of stillness.

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (1)

We will continue by presenting some notions necessary to the introduction of these theorems using the basic formula called Green-Riemann.

If MC [R.sup.2] is an elementary compact with the FrM border and P and Q are two functions of class [C.sub.1], on a simple convex domain which contain M, then

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (2)

By using the fundamental Green-Riemann's formula with P and Q chosen, from formula (1) we get:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (3)

By using the definition with Riemann's sums in (3) and then, for each point from the outline we can find out the ray by using the relation:

[R.sub.k] = [square root of [([x.sub.k] - [x.sub.k]).sup.2] + [([y.sub.k] - [y.sub.c]).sup.2] (4)

Talking into account the fact that fruits have usually unique shapes and thus, they can have sizes and shapes that are very different, the rays have been normalized with the medium ray and next by applying the Fourier transformed over the signature of the shape, we get:

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII] (5)

The histogram applied to the Fourier transformed shows how many times each of the values of the rays can appear.

4. SHAPE RECOGNITION BASED ON ARTIFICIAL NEURAL NETWORK

Consequently, as a result of tests and attempts carried out, a neural network was chosen, of the type back propagation having 16 neurons on the first stratum, 24 neurons on the second one, as for the third one, the number of neurons varies from 2 to 3 according the requirements.

Considering the fact that the algorithm is supposed to classify to be of hyperbolical tangent type for the first two strata and of sygnoidal type for the third stratum.

f (V) = 1 / 1 + [e.sup.-qv], f(v) - 1 - [e.sup.-2v] / 1 + [e.sup.-2v]; (6)

The exit stratum is formed by 2 or 3 neurons (Fig.2) that signify the two or three categories: according to the basic shape--good products (1,0,0), average products (0,1,0) and under standard products (0,0,1), according to the integrity of the products--good products (1,0), damaged products (0,1) and according to the size: big products (1,0,0), medium products (0,1,0) and small products (0,0,1).

The precision of the algorithms implemented was tested by using tomatoes and the network was trained up to the point when the average global square error reached the value of under 0,0001 during the training stage.

[FIGURE 2 OMITTED]

5. RESULTS AND DISCUSSION

After the training of the network, tests were carried out upon it while using 50 tomatoes belonging to all classes according to the size, quality or the presence of fault, and the evaluation of tomatoes belonging to each class was carried out by a human operator.

Very good results of the tests were obtained, with an average of the classification precision of over 94 % in the case of setting the integrity of products, 99 % in the case of evaluating the size and 91 % for the evaluation of quality.

We consider that the lower percent than 91 % in the case of products quality assessment is mainly due to the small amount of this training stage, as well as the inappropriate choice of products from the training set. Consequently, we can take it that the usage of neural networks in the determinating process is a fast, secure one, very easy to implement.

Also, it allows the assessment of a large number of features and it can be easily correlated to field of artificial intelligence such as the fuzzy logics reason.

6. CONCLUSION AND FUTURE WORK

After the test the following conclusions have been formulated:

The technique of the signature of rays is quite easy to apply, is quick and can be used both in the classification of the products according the shape and in the analysis of the products quality.

It can be easily correlated with the classification of neuronal type.

Taking into account the numberless amount of forms that the vegetal products can have, the utilization of the classificatory of neuronal type can be considered an important alternative (in comparison with the rest of the techniques of classification). The HIS pattern (used for image analysis) also proved efficient, saving a large amount of time and resources.

7. REFERENCES

Baoping, J., (1999). Non-destructive Technology for Fruits Grading. International Conference on Agricultural Engineering Beijing

Buzera, M.; Stefan C., (2007). Detecting the integrity of the shape of vegetal products by using non-destructive techniques, Proceedings of the International Conference, Bulgarian National Society of Agricultural Engineers, Lozenec, Vol.2 (2007) 252-256

Buzera, M.; Voicu, G.; Prostean, G. & Prostean, O. (2008) Techniques of Analysing the Colour of--roduces for Automatic Classification, 12th IEEE International Conference on Intelligent Engineering Systems, Miami, Florida, USA, 209-214

Heinemann, P. H.; Pathare, N. P. & Morrow, C.T. (1996), An Automated Inspection Station for machine vision grading of potatoes. Machine Vision and Applications 9. 14-19

Laykin, S.; Alchanatis, V. ; Fallick, E. & Edan, Y. (2002). Image--rocesing Algorithms for Tomato Classification. Transaction of the Assae. Vol. 45(3) 851-858

Laykin, S.; Edan, Y.; Alchanatis V. & Regev, R. (2003). Development of a Quality Sorting Machine Using Machine Vision and Impact. Transaction of the ASAE

Sudharkara, R. P.; Gopal, A.; Revathy, R. & Meenakshi K. (2003). Colour Analysis of Fruits Using Machine Vision System for Automatic Sorting and Grading. Journal Instrum. Soc. India. Vol.34(4) 284-291 Tao, Y. & Wen, Z. (1999). An adaptive spherical Image Transform for High Speed Fruit Defect Detection. Transaction of the ASSAE. Vol. 42(1) 241-246