New algorithms for color classification with machine vision devices.

Buzera, Marius ; Olaru, Onisifor ; Uscatescu, Mihai 等

1. INTRODUCTION

Over the last few years a large amount of algorithms assessing colour describers has been developed, yet most of them being hinted to be used in classifying industrial products, while their adjustment to the classification of vegetal products does not provide the best results. Thus, in 1999, Baoping presents a part of the identification and assessment algorithms of the colour describers, pointing out to their importance in the automatic classification processes. In 2002, Luzuriaga presents the disadvantages of measuring by colorimeters or spectrophotometers and he also develops an assessment algorithm of the colour of food items based on the analysis of areas. Still in 2002, Laykin develops an automatic classification devices for vegetal products on the basis of colour describer, all of that while valuing the mean and standard deviation of all pixels in the product area. Sudhakara in 2002, presents and validates a series of identification algorithms of colour by using neural networks. A quick method of orange classification was developed by Cruvinel in 2002, based on image processing with correlation analysis in frequency domain while Nordan in 2007 validates a series of classification algorithm of potatoes according to the shape, colour, and analysis of faults via classification system. Buzera in 2008 developed a series of identification algorithms of the shape, colour and size of the products by using processing techniques of images, fuzzy algorithms and neural networks, these algorithms being validated by carrying out an automatic classification device and by using machine vision techniques. A few results and conclusions to implementing and testing the mean and the standard deviation algorithm as well as the analysis of the colour histogram average, are presented throughout the article.

2. THE EXPERIMENTAL DEVICE AND THE ANALYSIS PROCESS DESCRIPTION

The device used during the tests, and presented in Fig.1, allows for both vegetal or industrial products classification having a shape factor close to 1, and using machine vision techniques.

The images are acquired through the intermede of two cameras CCD (CV1 and CV2) and transmitted to a calculus system (SC), which is to convey comends to the execution fields of the installation via a command circuit (CCM) of the engines, after taking out the shape describers and assessing them.

[FIGURE 1 OMITTED]

The device ensures the classification of the products into three classes but if new patterns are being added, the number of classes may increase by two. Two categories' of products have been used for the tests: balls and tomatoes. Thus, 30 tomatoes and 30 balls have been chosen, 10 of them being red, other ten green ones, 10 pink. Their affiliation to the 3 classes was carried out by a human operator, the products being placed by human hand and one by one on the roll conveyor. The first algorithm tested was the one analyzing hue histogram average, followed by that of determining mean and standard deviation, while the accuracy was hinted for the both of them.

3. PRESENTING THE ALGORITHMS

The two images acquired by the cameras are turned from RGB (red, green, blue) format to HIS (hue, intensity, saturation) format, and in order to point out the main features in classification, the hue histogram is to be used.

The tests carried out have pointed out have pointed out to the fact that this pattern introduces an considerable decrease of the calculation time by reducing the three dimensions of the RGB format to a single one: the H--hue and, more over, the hue becomes invariant to the light intensity alterations.

In order to divide the images so as to suitably identify the area of the products under analysis, a specific algorithm based on the identification of the optimum level on the basis of assessing the weight mass of the image, has been used.

The algorithm assessing the average of the histogram projected triggers the caring out of a segmentation of pixels belonging to the product, that for an identification of the analysis describers of colour, while the maximum value MAX, corresponding to the product, together with the value of the abscise vector X are to be extracted from the histogram vector obtained. In order to assess the homogeneity of colour, the average of the pixels values from the nearby of the maximum value identified is to be used, as one can easily notice in Fig.2. Thus, the X value points to the main colour of fruit, while MAX1 points to the frequency of its appearance.

[FIGURE 2 OMITTED]

Both X value and MAX1 value are determined by tests, thus setting the variation values for each class of products. Taking into consideration the high amount of parameters in view, the describers obtained are provided to a fuzzy decision block, developed on the basis of a function with trapezoidal affiliation. The mean and standard deviation algorithm is simply and suppose the calculation of the average and the standard deviation for all pixels belonging to the product from the image acquired, while the variation fields for the three classes have been identified through tests on similar products.

4. RESULTS

In order to test the algorithms proposed 30 tomatoes and 30 bills belonging, to the same type as those used during the training period have been chosen.

The main parameters hinted during the testing of the two algorithms were the precision of the classification as compared to the speed of the conveyor, as well as the classification speed. The algorithm based on the calculation of the average of the histogram provides very good result allowing for a quick a correct identification of variation field for each class as one can also see in Fig.3.

[FIGURE 3 OMITTED]

As one can notice, the max1 value was found and refound in the colour field identified during the training period. Thus, the tests carried out at various speeds of the conveyor, have allowed for setting an accuracy in the classification of tomatoes of 97%. As one can also see in table 1 the procent of 90% in the classification of pink tomatoes is mainly due to the choice of a too little training set of products. For balls, where the colour is much more homogeneous, the classification accuracy is 100%.

Also for the mean and standard deviation algorithm, the results are good, the accuracy is 90% for tomatoes and 100% for the balls, as one can notice from table 2.

5. CONCLUSIONS

The tests and experiences carried out have pointed out to the following conclusions:

--any of the algorithms are not influenced by the speed of the conveyor, thus providing results irregardless of its speed;

-the mean and the standard deviation algorithm imply a low amount of calculus, it is faster, but it doesn't provide enough information for the analysis of the hue homogeneity of vegetal product. Also, a high amount of hues of such a product depending on the environment may cause real problems to the product. On the contrary it is but very efficient in the analysis of the colours of the industrial products, because the homogeneity is a much higher one in this case.

--the calculation algorithm of the histogram average is rather easy to implement and it also provides good results, irregardless of the variation degree of the colours of the vegetal products from a set, but it depends a lot of the representativity of the set of products chosen for the training, in order to identify the variation fields of the colour for each class of products. Plus, this algorithm introduces an invariance to the sudden alteration of the values of the hue duet o the light modifications.

--the fuzzy technique provides much better results than the classical decision algorithms, especially if their amount and variation degree is a very high one.

6. REFERENCES

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

Buzera, M.; Prostean, G. & Stefan, C. (2008). Assesment techniques of the products on the basis of shape analysis , The 19th International DAAAM Symposium, Viena

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

Cruvinel, P. et al. (2002). Image Processing in Automated Pattern Classification of Oranges. Trans. of the ASAE

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

Luzuriaga, D.A. & Balaban, M. O. (2002). Color Machine Vision System: An Alternative for Color Measurement American, The World Congress of Computers in Agriculture and Natural Resources, Brazil, pages 93-100

Noordam, C. et al. (2000). High speed potato grading and quality inspection based on a color vision system, AGENG

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

Tab. 1. The variation of the classification accuracy
while using the calculus algorithm of the histogram
average

The speed The class The no. of The no. of The
of the products products accuracy
conveyor to be suitably
[m/s] tested tested

0,05--0,1 Red 10 10 100%
 tomatoes
 Pink 10 9 90%
 tomatoes
 Green 10 10 100%
 tomatoes
 Red balls 10 10 100%
 Pink balls 10 10 100%
 Green balls 10 10 100%

Tab. 2. The variation of the classification accuracy
while using the mean and standard deviation algorithm

The speed The class The no. The no. The
of the of balls of balls accuracy
conveyor to be suitably
[m/s] tested tested

0,05-0,1 Red 10 9 90%
 tomatoes
 Pink 10 8 80%
 tomatoes
 Green 10 10 100%
 tomatoes
 Red balls 10 10 100%
 Pink balls 10 10 100%
 Green balls 10 10 100%