An Automatic Sorting System For Recycling Beverage

Published Date: 02 Nov 2017

This article describes the prototype implementation of a real-time automatic identification and sorting system for recyclable beverage cans using an intelligent computer vision technique. The image recognition system was developed based on eigenface algorithm and achieved its ability to identify and sort by means of automatic learning process. Three experiments have been conducted based on pose, position and types of beverage cans moving on a conveyor belt. The results show that the identification and sorting of beverage cans registered an accuracy up to 95%. It can be concluded that the performance of the proposed system is robust.

Keywords: word; automatic sorting, eigenface, detection, beverage cans, pattern recognition

Introduction

Recycling activities in Malaysia are still at infancy stage. Public awareness on the importance of waste recycling is relatively low. The amount of daily waste generated continues to increase due to increasing population and consumption. Table 1 shows municipal waste generation trends in Malaysia, showing a significant increase from 18,494 tons per day in 2002 to 26,419 tons in 2007. Currently, over 30,000 tons of solid waste is produced each day in Malaysia and less than 5% of these are being recycled [1]. In 2002, the estimated average waste generation rate was 0.96 kg per capita per day [2, 3], which has escalated to an estimated to 1.7 kg/person/day in 2010. If the trend continues, the amount of total waste generation would be expected to rise to 50,000 tons by the year 2020 [1]. It is also estimated that the quantity of solid waste collection is less than 70 % of the total amount generation. The remainder is believed to be disposed in illegal dumps with the consequent problems of environmental pollution.

Table 1. Domestic waste generated daily for the states of Malaysia [1]

States

JICA study estimation: tonnes per day (TPD)

Projection

in TPD

2002

2004

2007

2010

Johor

2,154

2,636

3,071

3,579

Kedah

1,309

1,602

1,866

2,175

Kelantan

1,073

1,313

1,529

1,783

Melaka

504

617

719

838

Sembilan

891

848

988

1,149

Pahang

1,024

1,253

1,460

1,702

Perak

1,644

2,012

2,344

2,733

Perlis

165

202

236

275

Pinang

1,026

1,266

1,462

1,705

Selangor

3,293

4,031

4,695

5,473

Terengganu

733

898

1,038

1,219

Kuala Lumpur

1,088

1,332

1,551

1,808

Sarawak

1,674

2,058

2,387

2,783

Sabah

2,085

2,517

2,962

3,418

Total

18,494

22,638

26,419

30,794

Beverage cans, which are mostly aluminium, constitute one of the non-degradable components in municipal solid waste. Recycling beverage cans has a number of significant environmental benefits such as the reduction of landfill space and greenhouse gas emission. It’s widespread practice can save up to 95% of the energy used to make aluminium cans from virgin ore. Recycling 1 tonne of aluminium saves the equivalent in energy of 2,350 gallons of gasoline [4]. This is equivalent to the amount of electricity used by a typical home over a period of 10 years [4, 5].

Table 2 shows recyclable materials collected in the month of November 2009 in the Federal Terrotory of Putrajaya, Malaysia. Even though the beverage can component is relatively low in terms of percentage, i.e. 4.0%, the total amount of involved is 3583 kg [6] which is still huge. The recycling of such an amount of beverage cans can lead to an equivalent energy saving equivalent to about 8,420 barrels of oil. In 2004, only 51.2 percent of beverage cans were recycled in USA, for which the energy required for its manufacture is equivalent to about 15 million barrels of oil [4]. 

Table 2. Total of recyclable materials collected by Alam Flora in November 2009.

Materials

Collected (%)

collected (kg)

Paper

91.6

82,802

Plastic

3.7

3,339

Beverage cans

4.0

3,583

Glass and e-waste

0.7

595

Total

100

90,319

Sorting of solid waste components is a major step in any recycling activity. In developed countries such as the USA, Europe and Japan, manual sorting systems are still commonly utilised. Recently however, the development and improvement of automated sorting techniques have been actively carried out using more advanced technologies. For example, imaging technologies using infrared, near infrared, X-ray, and optical sensors have been deployed [7-9]. For recyclable plastic containers, sorting is generally based on resin type and colour characteristics [10].

The aim of the proposed study is to develop an automated identification and sorting system for recyclable metal containers that utilizes an intelligent machine vision system. The proposed system can provide an efficient and cost effective method for sorting recyclable metal containers, thus increasing recycling rates and contributing towards the saving of resources, mitigation of environmental pollution and reduction of landfilling space.

A computer vision system used in an automatic waste sorter basically captures and processes images based on object detection and recognition. There are a number of object detection and recognition methods that have developed, such as object detection and localization by dynamic template warping [11, 12], 3D object recognition on symmetry and virtual views [13], object detection with Gaussian distribution and clustering [14], object recognition based on image invariants [15], neural network-based face detection [16], training support vector machines (SVM) for face detection [17], smart sensing approach based on multi-resolution template matching [18], RFID and integrated communication technologies [***], and combined sharpening and edge detection techniques [19]. However, research on object detection with various background lighting conditions on the conveyor belt are few and not yielding significant output as yet.

The computer vision system developed in this work consists of three parts i.e. vision, recognition and action. The vision part is responsible for providing essential information to the computer system. The said information is an image, collected by a webcam integrated in the system. Data is transferred to the computer system after conversion into image formats. The recognition phase is responsible for making decisions based on perceived information i.e. whether it is a recyclable beverage can or otherwise. Once the detection is done, the system sends the appropriate commands to the action hardware apparatus for taking necessary sorting actions.

The composition of solid waste is unpredictable and different types of cans may exist in them. The different varieties of cans have different types of selection criteria. Since the objective is to recycle just beverage cans, it is sufficient to create selection rules for two categories, i.e. recyclable beverage cans and non recyclable cans. In this prototype system, the eigenface method is used.

Material and Methods

The main components of the automatic sorting prototype system for recyclable beverage cans are the webcam, PC, conveyor belt, push arm, and recycle bin. The conveyor belt drops the beverage cans into the bin while the Intel core2 2.4 GHz PC is used for image processing and vision based analysis. The webcam is placed vertically over a stretch of the conveyor belt for image acquisition.

Figure 1. Layout of the automatic sorting prototype system.

Figure 1 shows the prototype of the automatic sorting system. The conveyor belt runs at a variable speed of 15-20 cm/s and is loaded with beverage cans manually. For identification and sorting purposes, the possible position of the beverage cans can be located at 9 places on the conveyor belt as shown in Figure 2. The centre position, i.e. location 5, is the best position on the conveyor belt for image acquisition through the webcam. This is because the database was generated from this location in order to reduce the computation time for object recognition.

Figure 2. Possible positions of the object on the conveyor belt

The webcam grabs the image automatically by detecting the object movement as shown in Figure 3. Once the object is detected, the image data is sent to the PC for the identification step. Then the recognition algorithm identifies the aluminium beverage cans between consecutive inputs. The algorithms for object recognition and identification are explained as follows.

Figure 3. Images captured by web cam

Object Recognition Algorithms

The eigenface technique is one of object recognition algorithms based on the method of Principle Component Analysis (PCA) [20]. In this technique, the training image is represented by a flat vector, which is combined together to form a single matrix. The eigenfaces of each training image are then extracted and stored into a database. When an object is identified and grabbed, its eigenfaces are calculated and compared with those in the eigenface database. The database contains 500 images of beverage cans, which are divided into two groups i.e. aluminium and non-aluminium beverage cans.

Eigenfaces are sets of eigenvectors that are used for computer vision identification. To produce eigenfaces, digital images of an object are collected at the same lighting conditions and normalized. The normalized data is then processed at the same resolution (n x n) for all digital images. However, the rank of the covariance matrix is limited by the number of training images. For example, if there are N training examples, there will be at most N-1 eigenvectors with non-zero eigenvalues. If the numbers of training samples are smaller than the dimensionality of the images, the principal components can be computed more easily as follows.

Let T be the matrix of pre-processed training samples, where each row contains one mean-subtracted image. The eigenvector, S is calculated by the covariance matrix,

where TTT is a covariance matrix.

The eigenvector decomposition of S is given by

However, if TTT becomes a very large matrix, the eigenvalue decomposition can be written as,

Multiplying both sides of the equation (3) with TT, we obtain,

where, ui is an eigenvector of TTT and vi=TTui is an eigenvector of S.

If we have a training set of 500 images of 100 x 100 pixels, the matrix TTT is a 500 x 500 matrix, which is much more manageable than the 10000 x 10000 covariance matrix. However, if the resulting vectors vi are not normalized, an extra step is to be applied for normalization.

Identification Process

Figure 5 shows the block diagram of the methodological process for the identification system.

Identification

Web Cam

Image capture

Normalized Image

Matching Process

Database

Figure 4. Block diagram of the identification system

The details of the identification processes are as follows:

Once an object on the conveyor is detected by a webcam, the webcam automatically captures a picture and sends it to the PC. The size of the image is 320 x 240 pixels in bmp format.

The next step consists of using median filtering to reduce the lighting, noising problems and preserve edges. Each layer is filtered with a modified Sobel mask of size 7 x 7.

The image is then normalized, the edges extracted and then stored as a logical image. If the edges are not closed due to lack of contrast in the image, the system closes the missing points in order to get a better picture. To store the image in normalized form, the normalization conditions are as follows:

Beverage can is placed right in the middle on the conveyor belt.

Background of the conveyor belt is green.

Lighting for each position is of the same intensity.

Distance between the camera and can is 60 cm.

Webcam resolution used is 100 x 100 pixels.

Once the sharpening and edge extraction is done, the silhouette for each object is produced and a set of features are calculated at the identification stage.

500 aluminium beverage cans are used for creating the database for testing purposes.

The matching process is undertaken by comparing the value of eigenfaces with the values in the database to find the closest values.

The identification process is handled by the system, after an object is detected through the webcam automatically. The captured object features are then matched with data for the beverage cans in the database. Image capture results are automatically saved with variable image1 and undergo a matching process by calling the command line as follows.

ClosestImage=Val(bevcans.Identify(Image1)) (5)

where bevcans is the object of classBevcansRecogniser function, in which it identifies the properties of the objects. The classBevcansRecogniser calculates the distance value of the object and compared with the distance value of the database. The distance value is considered valid if it is found with the smallest distance value of the bevCanstemplate function.

To determine the accuracy of the identification system, the beverage can images are stored in the database with a standard pose. The standard pose of the beverage can is in the middle of the conveyor belt as shown in Figure 2, with adequate lighting.

The different positions for sample identification are shown in Figure 5. Changes of pose did not affect the accuracy of identification and the system successfully identified the different types of cans. Initial results suggested that the eigenface function can be used to differentiate aluminium cans from non-aluminium cans.

Figure 5. Different positions for sample identification.

Results and Discussion

In the prototype identification system, a QuickCam Vision Pro webcam was used with a maximum resolution of 960×720 pixels and up to 30 frames per second. However, the system used a resolution of 320 x 240 pixels because it is already sufficient to obtain a clear enough image for identification.

Testing was done under specific conditions to evaluate identification performance with different positions. The first experiment was performed using 211D aluminium cans to capture the nine positions on the conveyor belt. The second experiment was conducted at one position with 6 different poses for 5 types of beverage cans i.e. 211D, Slim 200D, standard 206D, 250ml round cans and 325 ml round cans. The third experiment was conducted at 4 positions with 6 different poses for 5 types of beverage cans.

Fig. 6 shows the first experimental results for beverage can identification based on conveyor belt positions. It was found that the changes of pose did not affect the accuracy of recognition, as the percentage of successful identification is consistently above 80%. Fig. 6 also shows that the best positions to identify the beverage cans are 2, 4, 5, 6, and 8, for which the identification accuracy rates are above 95%. Position 5, which is the standard pose at the middle of the conveyor belt, registered an accuracy rate of 98.5%.

Figure 6. First experimental results for beverage can identification.

The second experimental results for beverage can identification based on 6 different positions for 5 types of cans are shown in Fig. 7. It shows that the slim200D and 211D beverage can types registered identification accuracy rates up to 95 %. This is because these two types of beverage cans have different characteristics from all other types of beverage cans.

Figure 7. Second experimental results for beverage can identification based on position, pose and types.

Figure 8 shows that the level of recognition for slim 200D and 211D types of beverage cans are very accurate, registering up to 95% correct identification for 4 positions with 6 different poses for 5 types of beverage cans. This is also due to the very different characteristics of both types cans when compared with other beverage cans.

Figure 8. Third experimental results for beverage cans identification based on position, pose and types.

These results show that the prototype identification system using computer vision is effective in sorting aluminium beverage cans from non-aluminium beverage cans. The system achieves this ability by means of learning by training. The training is performed using an active learning method to minimise the amount of data used. Fig. 7 and Fig. 9 show that the accuracy of identification varies based on position, pose and types of beverage cans. This is due to the different characteristics of certain types of cans compared to others.

Conclusion

The experimental results show that the accuracy of the identification and sorting system are strongly dependent on the quality and quantity of the data used in training and lighting illumination. Overall, the sorting system is able to perform the classification of beverage cans with more than 90% accuracy. The change of pose of the beverage can does not significantly affect the recognition accuracy for object identification. It is found that positions 2, 4, 5, 6, and 8 on the conveyor belt are the best positions for identification of beverage cans. Position 5 registered the highest identification accuracy. Apart from can position, identification accuracy also depends on the quality and the quantity of the loaded data in the database in terms of length, width, area and perimeter of each object type. More advanced algorithms may be required to improve system performance, but early results suggested that the K-means clustering technique may provide unique characteristics and could be used to better categorise aluminium cans and non-aluminium cans.