PROBABILISTIC WHITE STRIP APPROACH TO PLASTIC BOTTLE SORTING SYSTEM

Mohd Asyraf Zulkifley, Mohd. Marzuki Mustafa, Aini Hussain

Department of Electrical, Electronic and Systems EngineeringFaculty of Engineering and Built Environment, Universiti Kebangsaan Malaysia

Bangi, 43600, Selangor, Malaysia

ABSTRACTOne of the most important steps in plastic recycling indus-try is waste sorting. Plastic wastes are usually sort into twomain categories, which are polyethylene terephthalate (PET)and non-PET. This paper proposes a probabilistic approachto automated plastic bottle sorting by integrating size, colourand distance modelling of the plastic waste. Firstly, whitestrips are identified by employing maximum likelihood ap-proach. Information on the white and grey strips is then an-alyzed by using maximum a posteriori method. Feature his-togram is built by factoring the output decision of each whitestrip with its size. Finally, likelihood test is performed to clas-sify the waste into PET and non-PET. Our algorithm performsthe best in all evaluation metrics compared to the benchmarkalgorithms. It is most suitable to be implemented in a factorywith the ever changing surroundings.

Index Terms— Maximum a posteriori, Likelihood test,Plastic recycling, Maximum likelihood classification, PETbottle identification.

1. INTRODUCTION

Efficient waste management is one of the most importantcomponents in building a sustainable community. Pollu-tion reduction and resource preservation can be improvedby recycling some key resources such as plastic, paper andaluminium. In Gulf countries alone, 120 million of tons ofwaste are produced in 2012 [1], where a majority of them areplastic materials. Plastic waste needs to be sorted first beforethey can be processed into plastic flakes. In plastic recyclingindustry, there two popular types of plastic that need to beclassified, which are polyethylene terephthalate (PET) andnon-Polyethylene terephthalate (non-PET). They need to beseparated because of the recycling techniques are differentand PET plastics are more profitable to the recyclable cen-ter. Most of the sorting centers employ low skill laborers toclassify the bottles manually. There are also some factoriesthat use some chemical substances to identify and sortingthe plastic wastes [2]. The downside to this approach is the

Thanks to Universiti Kebangsaan Malaysia for funding through grantGGPM-2012-062 and Suzaimah Ramli for the database.

chemical residues from the process, which still bring harmto the environment even though it is very minimal. To elimi-nate this risk, a sensor based system is proposed to automatethe sorting process such as the system by Picon et al. [3],where a CCD camera is utilized. In this paper, we propose aprobabilistic approach to plastic waste sorting through com-puter vision technique. Our main novelty is the detectionscheme that sorts the bottles probabilistically based on thedetected white regions and its neighbourhood data into PETand non-PET regions. We named the contour box of the con-tour sub-region as the white strip. A maximum a posterioriapproach is used to collect the decision of each sub-regionsinto a histogram before a likelihood test is performed to final-ize the sorting. We will utilize all three channels of the RGBcolour model instead of grey value only.

One of the earliest vision-based plastic bottle sorters isbased on near infrared camera [4]. They transformed thecaptured image into wavelet representation so that coefficientset of each type of plastics can be extracted. The classifi-cation method employs a simple Euclidean distance, whichis less accurate since some of the material’s property is verysimilar to each other. Tachawali et al. [5] then introduceda multi-stage classification method to improve the segmen-tation. Their method integrates quadratic discriminant andtree classifiers by assigning a weightage to each classifier forbetter accuracy. The system searches for any dip in the re-flectance values, where the ratio of the dips is used as the mainindicator. Their method requires the bottle to be in uprightposition since they sample the top and bottom of the bottle toremove the noise form the bottle’s label. They employed prin-cipal component analysis to perform the adjustment, whichadds another computational load to an already slow system.Instead of using near infrared information, Ramli et al. [6]employs grey level image to classify the bottle based on theintensity information. Label on the bottle is removed first, sothat regions of interest (ROI) that will be built are free fromthe label information. Pixel intensity of each layer of ROI arecollected into a histogram, which will be the input to the lin-ear discriminant analysis to sort the plastics. The performanceof their method will degrade if the bottle is not in upright con-dition and if the label area is not limited to the middle regiononly.

3162978-1-4799-2341-0/13/$31.00 ©2013 IEEE ICIP 2013

Fig. 1. Overview of the full sorting system.

Scavino et al. [7] focuses on separating several touch-ing plastic bottles on the conveyor line by using genetic algo-rithm. They assumed that a line can be used to separate twobottles, which indirectly indicates that no overlapping is al-lowed. Fifty lines are constructed between two bottles, whichare built based on the heuristic rules. Genetic algorithm isthen used to obtain the line with an optimal division. Varioustypes of kernel have been utilized by Shahbudin et al. [8] toquantify the angle of the detected bottle’s edge. This edge iscollected into a histogram, where it will be classified by usingsupport vector machine. Their assumption is a non-PET plas-tics is more likely to have an equal length of straight line be-tween width and height of the plastic bottle. However, it canbe misleading if the bottle is already bent or crumpled. Byusing similar classifier, system in [9] emphasizes on real timeapplication. Foreground subtraction is used to extract the bot-tles and region growth algorithm is applied to smooth out theforeground. Only histogram of the intensity is used so that thecomputational load will be low but will result in low accuracy.This paper is divided into four sections, which starts with theintroduction. Section II discusses in details on the proposedmethodology on probabilistic white strip approach. Simula-tion results and discussion are given in section III. Section IVconcludes the paper, and some future directions are proposed.

2. METHODOLOGY

The basis of our work is to make a decision based on theneighbourhood information of the white strip. Normally,white strip is detected because of the reflection at the corner-ing surface. For PET bottle, the neighbourhood colour of thewhite strip will mostly be grey color due to the transparentproperty. On contrary, non-PET bottle is more likely to beother colours than grey. Thus, our algorithm is built by as-suming that for any detected white strip, the neighbourhoodcolor information that is parallel with the strip’s length shouldbe mostly grey. The full flow chart of the system is shown in

Fig. 2. An example of the processed image. (a) Input image(b) Extracted foreground (c) White strips (d) Contour boxes.

Figure 1.Firstly, the video images (Figure 2(a)) of the bottles on

the conveyor belt are captured. Then, foreground mask (Fig-ure 2(b)) is built to differentiate the bottles and the conveyorbelt. Basic subtraction method is used and connected pixelsof the foreground are grouped together. Minimum numberof connected pixel is observed before performing erosion anddilation filters. White strips (Figure 2(c)) are then checkedfor the entire image by assuming 3D exponential distributionof the RGB channels. White strip likelihood, L1 is assumedif the pixel values are close to 255 for a 24-bit RGB image.On the other hand, likelihood, L2 for not a white strip is alsomodelled by exponential distribution, but bias towards zerovalues. Each connected pixel for that particular sub-regionwill be integrated before maximum likelihood is used to iden-tify either it is a white strip or not, which is indicated as D2.

L1 =1

Z1exp−λ1X1 (1)

L2 =1

Z2exp−λ2X2 (2)

D1 = arg maxi∈{1,2}

(Li) (3)

where X1 and X2 are the difference in RGB values with re-spect to the white and black pixels respectively, while Z is thenormalizing factor. A contour box (Figure 2(d)) is built foreach detected white strip, W and 8-neighbourhood box willbe built around that strip as shown in Figure 3. Maximuma posteriori is employed to make decision D2, either the de-tected white strip has occurred because of the reflection sur-face or not. Based on our earlier assumption, a grey neigh-bourhood or a grey strip, g indicates that the likelihood of theplastic to be PET is high. Let Y be the observation input andµ1 be the vector of average pixel value of R, G and B chan-nels.

D2 = arg max(p(g|Y )) ∝ p(Y |g)p(g) (4)

3163

Fig. 3. 8-connected box for each white strip.

Two priors and one conditional probability are used to modelthe grey strip. Fist prior, p(D) is based on exponential dis-tribution of the distance d that represents the length betweencenter of the anchored contour box to the 8-connected box.The hindsight behind this assumption is the occurrence of agrey strip will likely occur close to the white strip for the PETmaterial. The second prior, p(R) concerns on the ratio be-tween the height (h) and width (w) of the contour box, whichwill be capped by threshold τ . Usually, a slender white stripwill have a higher probability that it composes of PET mate-rial.

p(g) = p(D)p(R) (5)

p(D) =1

Z3exp−λ3d (6)

P (R) =

1 if h > τw

1

Z4exp−λ4

hw if h ≤ τw

(7)

The likelihood is modelled by a 3D Gaussian distribution,where the mean is the difference between the observationsand µ1. The covariance matrix, Σ is a diagonal matrix withequal values (ΣR = ΣG = ΣB). µ1 is selected based on as-sumption that a pixel will likely be grey colour if each valueof the RGB channel for that particular pixel is close to theiraverage value.

p(Y |g) = 1

Z3exp−(Y−µ1)

TΣ(Y−µ1) (8)

The highest probability will be chosen as the representative todecide either the white strip has a grey neighbor or not. Then,a histogram (H) will be built based on D2 where a weightagewill be integrated that depends on the white strip size. H1 isa histogram representing PET and H2 represents non-PET.Lastly, likelihood test is performed to distinguish between

PET and non-PET plastic. Both PET and non-PET likeli-hoods are modelled by Gaussian distribution, where theirmeans (µ2,PET&µ2,non-PET) and variances (σ2,PET&σ2,non-PET)are obtained through supervised training. Final decision onthe material type is denoted by D3.

D3 =

{PET if L3 > L4

non-PET if L3 ≤ L4

(9)

L3 =1

Z4exp

−(

(H1−µ2,PET)2

σ2,PET

)(10)

L4 =1

Z5exp

−(

(H2−µ2,non-PET)2

σ2,non-PET

)(11)

where L3 and L4 represent the likelihood of PET and non-PET plastic respectively. Then, a pneumatic mechanism willpush the bottle to the right collection bin for further recyclingprocess.

3. SIMULATION RESULTS AND DISCUSSION

300 images of plastic bottle waste have been put on the con-veyor belt to validate the accuracy of our system. 106 of thetotal bottles are made from PET while the rests are non-PET.The tested bottles are of various sizes and colours. The labelon the bottle’s surface has not been peeled off to simulate realapplication in the industry. Some of the bottles have dentedsurfaces, which add the difficulty to distinguish the materials.The simulation videos also consist of lighting change scenesto simulate the robustness of the algorithm to sudden andglobal illumination variations. Figure 4 depicts some samplesof PET and non-PET bottles where darker environment can beobserved in certain frames. The frame size is 320 × 240 andall tested algorithms are able to process at least three framesper second on Intel Core2 Duo 2.4 GHz machine. Four per-formance metrics are used, which are accuracy (A), error (E),precision (P) and recall (R). Let Tn be the total number ofthe test bottles, tp be the true positive detection, fp be thefalse positive detection, tn be the true negative detection andfn be the false negative detection.

A =tp + tnTn

(12)

E =fp + fn

Tn(13)

P =tp

tp + fp(14)

R =tp

tp + fn(15)

Both algorithms by House et al. [9] and Shahbudin et al. [8]are selected as the benchmarks for performance comparisonto our algorithm, which will be indicated as Zulkifley et al.Table 1 shows the performance of all algorithms. In general,

3164

Fig. 4. Input samples: (a) non-PET bottles (b) PET bottles.

Table 1. Performance comparison between algorithms byZulkifley et al., Tachawali et al. and Shahbudin et al.

Algorithm A E P R

Zulkifley et al. 0.807 0.193 0.709 0.750Shahbudin et al. 0.623 0.377 0.320 0.077House et al. 0.400 0.600 0.096 0.087

the algorithm by Zulkifley et al. performs the best, followedby Shahbudin et al. and House et al. The accuracy of ouralgorithm is 0.807, which is more than twice compared tothe algorithm by House et al. The main reason is our algo-rithm utilizes more robust detection method, especially dur-ing illumination changes by integrating all three channels ofRGB instead of grey channel only. Besides, our probabilis-tic approach is able to adapt to global lighting changes sinceall hypotheses will be adjusted accordingly when illuminationchange occurred.

Method by Zulkifley et al. also records the best precisionand recall values of 0.709 and 0.705 respectively. On con-trary, the recall values for both methods by Shahbudin et al.and House et al. are less than 0.1. However, House et al.’salgorithm performs better in precision evaluation as it records0.377 compared to 0.096 for House et al. This weakness canbe attributed to modelling by a single parameter, as opposedto hybrid modelling of distance, size and colour informationfor our probabilistic approach. Multiple cue detections areneeded because of the physical similarity between PET andnon-PET. Besides, our grey strip detection is robust to rota-tional changes where the transparent surface can be detectedwithout the needs to rotate the bottle to the upright positions.Zulkifley et al.’s algorithm also mitigates the effect of the la-bel by assigning a prior to the ratio of length and width ofthe white strip. This will reduce the probability of PET de-tection as PET bottle is more likely to have a squarish shapeinstead of a slender box. Receiver operating characteristic(ROC) graphs for all tested algorithms are depicted in Figure5. The graph proved that our algorithm is the most robustwith the biggest area under the curve, followed by Shahbudin

Fig. 5. Receiver operating characteristic for the algorithms byZulkifley et al., House et al. and Shahbudin et al.

et al. The performance of our algorithm is stable over a widerparameter range, which will allow the system to perform wellin various operating conditions. Algorithm by House et al.is the least robust where area under the curve is the small-est, which complicates parameter tuning process because ofnarrow optimal operating window.

4. CONCLUSION

In conclusion, the algorithms by Zulkifley et al. performs thebest with the highest accuracy of 0.807, followed by Shabudinet al. and House et al. with 0.623 and 0.400 respectively.We achieved a good performance because of integrated mod-elling of the physical appearance of the bottle. Moreover, ouralgorithm is the least affected by rotational and illuminationchanges. The main novelty of our approach is probabilisticrepresentations of the white and grey strips that allow us toproduce more distinctive features between PET and non-PETplastics. The algorithm can be improved by considering par-allel programming for faster implementation and colour con-stancy [10] for better accuracy under light changes.

3165

5. REFERENCES

[1] M. Al-Maaded, N.K. Madi, Ramazan Kahraman,A. Hodzic, and N.G. Ozerkan, “An overview of solidwaste management and plastic recycling in qatar,” Jour-nal of Polymers and the Environment, vol. 20, pp. 186–194, 2012.

[2] L. Bartolome, M. Imran, B. G. Cho, W. A. Al-Masry,and D. H. Kim, Recent Developments in the ChemicalRecycling of PET, chapter 2, pp. 65–84, InTech, 2012.

[3] A. Picon, O. Ghita, P.F. Whelan, and P.M. Iriondo,“Fuzzy spectral and spatial feature integration for clas-sification of nonferrous materials in hyperspectral data,”Industrial Informatics, IEEE Transactions on, vol. 5, no.4, pp. 483 –494, nov. 2009.

[4] J. M. Barcala, J. L. Fernndez, J. Alberdi, J. Jimnez, J. C.Lzaro, J. J. Navarrete, and J. C. Oller, “Identification ofplastics using wavelets and quaternion numbers,” Mea-surement Science and Technology, vol. 15(2), pp. 371–376, 2004.

[5] Y. Tachwali, Y. Al-Assaf, and A.R. Al-Ali, “Automaticmultistage classification system for plastic bottles recy-cling,” Resources, Conservation and Recycling, vol. 52,no. 2, pp. 266 – 285, 2007.

[6] S. Ramli, M. M. Mustafa, A. Hussain, and D. A. Wahab,“Histogram of intensity feature extraction for automaticplastic bottle recycling system using machine vision,”American Journal of Environmental Sciences, vol. 4(6),pp. 583–588, 2008.

[7] E. Scavino, D. A. Wahab, H. Basri, M. M. Mustafa, andA. Hussain, “A genetic algorithm for the segmentationof known touching objects,” Journal of Computer Sci-ence, vol. 5(10), pp. 711–716, 2009.

[8] S. Shahbudin, A. Hussain, D. A. Wahab, M. M.Marzuki, and S. Ramli, “Support vector machines forautomated classification of plastic bottles,” in 6th Inter-national Colloquium on Signal Processing and Its Ap-plications (CSPA), may 2010, pp. 1 –5.

[9] B.W. House, D.W. Capson, and D.C. Schuurman, “To-wards real-time sorting of recyclable goods using sup-port vector machines,” in Sustainable Systems and Tech-nology (ISSST), 2011 IEEE International Symposiumon, may 2011, pp. 1 –6.

[10] M. A. Zulkifley, B. Moran, and D. Rawlinson, “Robustforeground detection: A fusion of masked greyworld,probabilistic gradient information and extended condi-tional random field approach.,” Sensors, vol. 12(5), pp.5623–5649, 2012.

3166