Quality assurance of refuse derived fuel with machine vision

(1)

Intelligent computing

Master’s Thesis

Pauli Immonen

QUALITY ASSURANCE OF REFUSE DERIVED FUEL WITH MACHINE VISION

Examiners: Associate Professor Arto Kaarna MSc. Keijo Manninen

Supervisor: Associate Professor Arto Kaarna

(2)

Lappeenranta University of Technology LUT School of Business and Management Intelligent computing

Pauli Immonen

Quality assurance of refuse derived fuel with machine vision

Master’s Thesis

2015

48 pages, 28 figures, and 3 tables.

Examiners: Associate Professor Arto Kaarna MSc. Keijo Manninen

Keywords: refuse derived fuel, image processing, machine vision, quality control, waste- to-energy, solid waste

This thesis studies the use of machine vision in RDF quality assurance and manufacturing.

Currently machine vision is used in recycling and material detection and some commercial products are available in the market. In this thesis an on-line machine vision system is proposed for characterizing particle size.

The proposed machine vision system is based on the mapping between image segmentation and the ground truth of the particle size. The results shows that the implementation of such machine vision system is feasible.

(3)

Lappeenrannan teknillinen yliopisto LUT School of Business and Management Älykäs laskenta

Pauli Immonen

Jätepolttoaineen laadunvalvonta konenäön avulla

Diplomityö

2015

48 sivua, 28 kuvaa ja 3 taulukkoa .

Tarkastajat: Associate Professor Arto Kaarna MSc. Keijo Manninen

Hakusanat: jätepolttoaine, kuvankäsittely, konenäkö, laadunvalvonta, energiajäte, yhdys- kuntajäte

Keywords: refuse derived fuel, image processing, machine vision, quality control, waste- to-energy, solid waste

Tämä diplomityö tutkii konenäön käyttömahdollisuuksia jätepolttoaineen laadunvalvonta ja valmistus prosessissa. Tällä hetkellä konenäköä on käytössä materiaalin tunnistamiseen ja kierrätyksen parantamisessa ja joitakin kaupallisia tuotteita löytyy markkinoilta. Lisäk- si suoraan kytketty konenäköjärjestelmä on esitetty palakoon jätepolttoaineen palakoon tunnistamiseen. Ehdotettu jätepolttoaineen palakoon tunnistus perustuu kuvassa segmen- toitujen palojen ja todellisen palakoon väliseen yhteyteen. Tulosten perusteella ehdotettu konenäköjärjestelmä on toteutuskelpoinen.

(4)

I wish to thank my supervisor and examiner Associate Professor Arto Kaarna and examiner Keijo Manninen. I would also like to thank Hannu Lepomaki and Ilkka Haavisto for arraging visits to refuse derived fuel manufacturing plants.

Lappeenranta, July 19th, 2015

Pauli Immonen

(5)

ABBREVIATIONS AND SYMBOLS

CEN European committee for standardization CDF cumulative distribution function

CDF cumulative distribution function ECD empirical cumulative distribution ECS Eddy current separator

EU European Union ICC intraclass correlation NIR near infrared

MSW municipal solid waste PSD particle size distribution PVC polyvinyl chloride RDF refuse derived fuel

SLIC simple linear iterative clustering SRF solid recovered fuel

SRM statistical region merging VIS visible spectrum

(8)

1 INTRODUCTION

1.1 Background

In the last decades rising energy cost and steady rise of crude oil price has made utilization of waste as fuel attractive for industrial enterprises. Other contributing factor is reluctance to dispose untreated waste in landfill sites. At the same time have material re- source consumption has increased and consequently, so has the amount of waste released to environment. A Vast range of the released waste may be utilized as refuse derived fuel (RDF) or solid recovered fuel (SRF). RDF or SRF is produced by shredding, dehydrating and removing non-combustible materials from municipal solid waste(MSW), industrial waste or commercial waste. SRF can be distinguish from RDF by the fact that it needs to meet the quality standards defined in CEN/TC 343[1]. The The most concerning problem in producing RDF or SRF is to produce it with steady quality due to heterogeneous nature of MSW. The most important quality attributes of RDF are caloric value, few impurities, low metal content and low chlorine content, defined particle size[2]. Particle size needs to be in certain limits or it may be defined with distribution. The particle size distribution depends on the application where the RDF is used.

1.2 Objectives and Restrictions

The objective of this thesis is to consider utilization of machine vision in quality assurance of RDF. The hypothesis is that machine vision system can produce useful information on RDF quality. The research objectives of this study are:

• Study the state-of-the-art quality assurance in RDF

• Study the use of machine vision in RDF quality assurance

• Investigate the use of sensor fusion to improve RDF quality

• Particle size detection

• Proposition of quality assurance system

Methods for detecting particle size with machine vision system is studied in the literature and an own implementation will be made. State of the art methodology for incombustible

(9)

objects and for recycling improvements is studied and the current solutions are presented.

1.3 Structure of the Thesis

This thesis is organized in 9 chapters. In chapter 2 the quality assurance and quality requirements of RDF is described. In Chapter 3 the manufacturing process and process steps are presented. In chapter 4 the sensory used in quality assurance and their fusion is described. In chapter 5 the use of machine vision in RDF manufacturing process is analysed and some state of the art machine vision solutions presented. In chapter 6 machine vision methods suitable for particle size detection are presented. In chapter 7 the experiments done during the project is bring forth. In chapter 8 the discussion about the research is given and future research directions are considered. Finally in chapter 9 the conclusions are drawn.

(10)

2 QUALITY ASSURANCE OF REFUSE DERIVED FUEL

2.1 RDF as energy

Combustion of waste derived fuels is currently being advocated. The use of such fuels as energy can mitigate greenhouse gas emissions and it can be economically advantageous for energy users. The biomass fraction of such fuels is considered to be carbon-neutral since the coal-dioxide is liberated from the combustion is recycled in plants. RDF processing facilities are usually near source of MSW. The combustion facility can be either near the source of RDF, integrated to the end of process line or it can be at remote location. RDF can be used as primary burner fuel or it can be co-combusted alongside with traditional sources of fuel such as wood chips or coal. The most used incineration technologies that are applicable for RDF are co-combustion, combustion in fluidized bed and gasification. The main advantages of using RDF as a fuel are important reduction in the volume of waste and the possibility of energy recovery[3]. In Europe the CEN pub- lished the standards EN 15359(2011) that establishes technical specifications for SRF. In the regulation SRF is defined as combustible obtained from non hazardous waste. The composition of municipal solid waste in different regions in Finland is shown in Table 1 [4]. According to this regulation SRF can not be produced from MSW due the content of hazardous waste. The amount of MSW produced in EU in 2004 was 259 million tons[5].

Figure 1 shows the general trend where MSW is diverted from landfills to recycling and waste-to-energy solutions in EU. From 1998 land the percentage of MSW disposed to landfill has decreased continuously in EU. Material recycling, composting and waste incineration has become more and more general in waste management.

Table 1.The composition of MSW in percents in different regions in Finland.

Place Metals Biowaste Glass Paper and

cardboard

Combustible waste

Non- combustible waste

Hazardous waste and E-waste South Kare-

lia

23.9 3.8 2.5 14.9 42.8 10.4 1.7

Turku 38.0 2.0 3.0 17.0 31.0 7.0 2.0

Päijät- Häme

23.0 5.0 3.0 12.0 32.0 21.1 6.0

Savonlinna 9.0 4.0 2.0 12.0 33.0 38.0 3.0

Helsinki area

39.7 3.4 3.9 16.9 32.5 2.6 1.1

Kuopio 35.0 3.0 2.0 11.0 23.0 26.0 2.0

Mikkeli 30.4 3.0 2.5 15.2 21.0 26.6 1.5

Kainuu 9.7 5.3 2.5 15.2 21.1 45.2 0.9

(11)

Inineration Material Recycling and composting Landfill

70 60 50 40

Percent of Total Waste Treatment

30 20 10

0

1996 1998 2000 2002 2004 2006 2008 2010

Figure 1.Waste management in EU in recent years.[6]

2.2 Quality criteria of waste fuels

Quality standards for SRF are partly of a legal nature. RDF incineration plant requires consistent RDF quality to ensure smooth incineration process. RDF must fulfill general quality requirements for safe and efficient utilization. Quality criteria and characterization of SRF in Finland is described in SFS-EN 15358:en by Finnish Standards association. The most important quality factors in the quality standardization are:

• Predefined caloric value

• Defined particle size and density

• Few impurities

• Low chlorine content

• Low heavy metal content

High chlorine content causes erosion in the boiler and the particle size has great effect on the application field in industrial utilization and the net caloric value. Operators usually

(12)

specify target values for chlorine, calorific value, moisture content and particle size. Waste incineration has a negative impact on air and water which can turn lead to environmental issues and detrimental health effects for humans. When incinerated the chlorine content of waste can cause corrosion and lead to hydrochloric acid and the release of highly toxic dioxins (PCDD/F). In order to minimize this risk European Union adopted directive [7]on the incineration of waste which was integrated into the directive [8] on industrial emissions. Different calculative methods suggest that 67% - 72 % of chlorine is originated from polyvinyl chloride (PVC) [9]. The quality of waste fuel is controlled by sampling.

2.3 Sampling

The objective of sampling is to characterize waste fuel by extracting a portion of it that is representative for the total waste volume[10]. The heterogeneous nature of MSW causes also the sampling results to be heterogeneous. That for the minimum number of independent samples will be 24 [11]. Analysis methods are used to define RDF sample properties. These include calorimeter combustion, Wickbold combustion X-ray fluorescence analysis[10]. The result characterization of the sampling is not on-line information, but characterization of the total waste volume.

2.4 Characterization of waste fuels

Apart from the legal requirements, additional requirements of chemical and physical parameters are usually laid down in the contract between the RDF supplier and user. Re- quirements may contain such parameters as particle size, net caloric value, chlorine content, moisture contend, bulk density and restriction of heavy metal content. The characterization is based on sufficiently extensive and accurate sampling, sample handling and standardised fuel analyses. In order to create a comparable framework of rules within Europe, the European Standardization committee (CEN) Founded Technical Committee CEN/TC 343 ’Solid recovered Fuels’ in 2002. Quality assurance is of SRF are based of the guidelines of the commitee [2].

(13)

3 RDF PRODUCTION

This chapter describes describes typical manufacturing process of RDF. The most common processing steps and equipment are presented.

3.1 Typical steps of RDF production

For high quality RDF multi-stage separation process is necessary. Depending on the input material the RDF production process will usually have the following steps[11]:

• Pre-crushing

• Mechanical Screening

• Metal Separation

• Post-crushing

• Confection

Pre-crushing is done by coarse sized crushers followed by mechanical screening with wind shifter or ballistic separation. Metal separation is magnetic separation for ferrous metals and Eddy current for non-ferrous metals. Confection phase consists magnetic separation and disc-screening. Disc screening is a method for separation by particle size.

Figure 2 shows an example process scheme of preparation of SRF. The SRF is used as primary burner fuel in a cement plant. In some cases like shown in Figure 2 separation of two-dimension and three-dimensional material is advisable. The 3D material usually contains more in purities like metals, stones and concrete. Therefore the 3D waste stream undergoes more complex procedure before it is shredded to its final particle size. For getting efficient metal separation performance magnetic and eddy current separation is done after each size reduction. Iron tends to be embedded in fluffy waste material and the metal separation is much easier to do when the structure of the waste is broken down.

(14)

INPUT

Commercial waste Packaging waste

Industrial waste Preprocessed Household

waste

Coarse sized crusher

Wind-sifter Air Classifier Magnetic

separator

Eddy-Current Separator

Wind Shifter

Fine size shredder

Middle size shredder Semi Product

storehouse

3D 2D

Magnetic separator (drum type)

Disc screen (<30 mm)

Magnetic separator (belt-type)

SRF

Metal recovery

oversize >

30mm Heavyweight

fraction (stones, concrete,etc)

Figure 2.Display of example process scheme for high calorific SRF [11].

(15)

3.2 Size reduction

Size reduction is commonly the first operation in RDF/SRF manufacturing process. Sev- eral process steps after size reduction require certain particle size to perform efficiently.

Secondary shredding is employed to adjust the final product size. The shrear shredder uses counter-rotating wheels having blades that shear the material as they are caught in the gap between the wheels [6]. Hydraulic drives are used, so that when jams occur, the rotor can be stopped.

Waste material input

In-feed conveyor

Cutter Shaft Cutter teeth

Discharge conveyor

Size reduced output

(a)

Size reduced output In-feed conveyor

Shaft Rotor

Grate with openings Hammer

Waste material input

Adjustable breaker bar

(a) Discharge conveyor

(b)

Figure 3.Display of size reduction principle of (a) rotary shear shredder (b) hammermill shredder [6].

3.3 Ballistic separation

Ballistic separation has a wider range of applications. Ballistic separator consist of rotating longitudes paddles that have a rotating offset against each other. Ballistic separator classifies particles to 3 different classes. Ballistic separation is illustrated in Figure 4.

Firstly the rotating movement causes low gravity flat shaped object(2-D) are moved up- wards in a circular movement by the paddles of the ballistic separator [12]. Secondly, 3-D material, stones hard plastics are bouncing or rolling downwards in a diagonal reverse direction. Lastly the perforation of the paddles allows fine fraction material such as debris, sand or dust to be screened.

(16)

Metal plate Heavy fraction

Waste objects

Rotating metal conveyors

Light fraction

Screen fraction

Crankshaft

Figure 4. Ballistic separation. [12]

3.4 Mechanical screening

The primary methods for screening are rotating screening or trommel screening and they are widely used in municipal waste processing. A variety of screens are used in RDF processing. Screening is utilized to separate oversized material after crushing or classify material flow according to particle size. Screens typically separate materials of smaller size from the mass passing over the screen. Disc screen, Figure 5 is one example where where smaller sized material drop down between the discs and bigger flows with the rolling discs. The screened particle size is determined by the shape and the area of the hole.

3.5 Air classification

Air classification is an important unit operation of separating particles according to their combination of size, shape and density[13]. The principle of the separation is based on different forces caused by air friction. Particles experience gravity and drag forces acting in opposite directions. Heavy particles, having terminal settling velocity larger that the velocity of air move downwards against the air stream, while light particles whose terminal settling velocity is smaller than the velocity of air rise along with the air stream to the top. Air classifiers can be grouped as static classifiers and dynamic classifiers. Static classifiers have no moving parts and the target size is adjusted by changing the direction

(17)

Material infeed

Oversized T

Accepts

″ Figure 5.Disc screen principle. [12].

and magnitude of the air flow[14]. Dynamic air classifier technology typically has a rotating plate on which the material is poured and dispersed with the aid of centrifugal force. Air classifiers are use in RDF process to remove debris from. Figure 6 shows the working principle of different static air classifiers. Air flowing from below scavenges and the particles are fed from above in cross-current direction in zigzag and shelf classifier.

Separation occurs in a cross-flow process. In horizontal classifier the

Figure 6.Static air classifiers (a) zigzag (b) shelf classifier (c) horizontal scavenging [13].

3.6 Magnetic separation

Magnetic separation attracts ferrous material and conveys it away. To work efficiently the material size has to be reduced to size between 10 mm and 100 mm. Overbelt mag-

(18)

netic separation is the most used magnetic separation technique in solid waste processing.

Overbelt separator is placed crosswise or lengthwise over the conveyor belt 7. From the flowing material iron objects are captured and carried away. As the iron objects leave the magnetic field they are dropped to appropriate canals or containers.

2

Figure 7. Overbelt magnetic separator [6].

3.7 Eddy current separation

Eddy current separation (ECS) is a widely applied method for separating nonferrous metals from mixed waste streams [15]. ECS uses a rotating drum with permanent magnets or electromagnet. ECS principle is based on a repulsive magnetic force between the varying magnetic field and a non-ferrous conductive particle. Lorenz force lifts up the non-ferrous metals and throws them in a longer trajectory than magnetic metals or non conductive materials. The Non conductive materials are only subjects to gravity force.

Particles are separated by placing a divider plate at a distance from the drum. The varia- tions of particle size shape, density, ballistic properties and conductivity describes a wide fan of trajectories[16]. It has a negative effect on the performance of the separation.

(19)

Figure 8.Non-ferrous metal separation with eddy current [16].

(20)

4 SENSOR FUSION

Sensor fusion is a technology that refers to the combination of sensory data from multiple sources to get more accurate inference than only from one sensor separately. Different sensors have their own advantages and limitations and perceptive uncertainties and sensor fusion is expected to reduce overall uncertainty and increase accuracy. Traditionally sensor fusion is defined in three different level. Data fusion, feature fusion and decision fusion [17]. [18].

4.1 Low-level fusion

The sensory data fused can be single or multidimensional signal. The signal fusion can be done in real time but due to the different sampling properties of multiple sensor the data need to be synchronized before the fusion. Commonly used fusion methods in pixel- level include Principal component analysis (PCA), wavelet transform fusion and band- rationing or combination of these methods[17]. Spectral detection is one example of pixel fusion and it is discussed in section 5.3.2.

4.2 Medium-level fusion

The data fused at medium level is features obtained from signals or images. Data fusion in this level is done by conceiting features of each sensor. Feature fusion is usually done in three steps, feature uniformization and normalization, feature reduction and concatena- tion, and feature matching. Classification methods are suitable for feature fusion. Least discriminating features might reduce the performance of most classification methods and some classification methods needs feature normalization to perform well[17].

4.3 High-level fusion

Unlike the low level fusion which deals the raw data from the sensory the information is threated after high-level symbolic presentation. The fusion process is to combine symbols associated with uncertainty measure to generate composite decision. High level fusion is also referred as decision fusion. In general high level fusion extract local decision from

(21)

each sensor and makes a combined decision out of them[17].

4.4 Sensories used in RDF process

For successful sensor fusion the sensory responses is expected to be independent from each others. In solid fuel processing there are used optical sensors, visible spectrum(VIS), near infrared (NIR) and X-ray for material sorting. Also, electro magnetic sensor has been used[19]. Optical camera estimates the reflection rate of the visible light and NIR camera the reflection rate of the near infrared light. X-ray gives information about of average atomic number and thickness of the material while electro magnetic sensor measures conductivity. In theprevious research mostly low-level fusion is applied. ZenRobotics is using multiple sensor on construction demolition waste sorting.[20] They are fusing, VIS, NIR, height map and metal detector data. Fusion is done in pixel level. There are examples how sensor fusion is utilized in section 5.3

(22)

5 MACHINE VISION AND RDF PROCESS

5.1 Introduction to machine vision

Machine vision methods have provided tools in industrial applications that require automatic inspection and product analyses. The application of machine vision methods enable collection of numerical information from industrial process. Machine vision process typically have the following steps [21]:

• Image acquisition

• Image preprocessing

• Segmentation

• Representation and description

• Classification or parameter estimation

Image acquisition can be done somewhere in the RDF production depending on the detectable feature. If the detectable feature is the particle size of the end product , the image acquisition should be done after fine shredding. If the detectable feature is analysis of input material composition the image acquisition should be considered before separation.

Machine vision can be also considered to be in flow of rejects from magnetic or eddy current separation to monitor the performance of recycling. Image preprocessing stage usually have noise filtering with median, averaging or Gaussian filtering. Also, gradient operators for edge enchantment are pre-processing steps in certain applications. Machine vision typically segments the informative objects in the image. The segmented parts are characterized by set of features.

5.2 Benefits of machine vision in RDF production

Machine vision can help to adjust automation of the SRF process online. Online particle size detection is able to detect worn blades in crushing or damage in mechanical screens.

The main indicator of normal crushing operation is the average size of particles. Particle size and net caloric value has shown to have strong influence to the utilization of the waste fuel[2].

(23)

5.3 Existing machine vision solutions

Overall, there appears to be scientific articles dealing with the area of visual sorting of incombustible and recyclable objects or particle size analysis of RDF[22], [22],[23],[24]

. However, in the field of machine vision and pattern recognition there are more general papers of visual sorting and segmentation that can be applied to RDF production. Parti- cle size analysis has been done in by particle image analysis method in[25] and for the proposed particle detection system samples needs to be prepared and are not for online measurement. The method also assumes that particles are not over lapping and are fully separable from the background.

5.3.1 NIR detection for incombustible particles

NIR spectroscopy has been used for non-destructive analysis of materials. Today many instrumentation and automation companies are selling dedicated NIR instruments for online process monitoring in numerous industry applications[23]. These instruments record sample transmittance or reflectance of set of wavelengths.

5.3.2 Hyperspectral detection for incombustible particles

Spectral imaging is becoming increasingly interesting not only for remote sensing use but also for industrial applications[26],[27]. Hyperspectral imaging is used to separated different materials from each others. Hyperspectral imaging technique represents an attractive solution for characterization, classification and quality control of different materials.

Studies have been carried out in solid waste recycling, with reference to paper glass and plastics[27].. Figure 9 shows the result of using HSI on material detection. In the figure 2 factors are excluded from spectral imaging to characterize materials. The eclipses shows confidence interval where 95 percent of samples can be found and which do not cross each other. Different materials can be well modeled with PCA

5.3.3 Machine vision products

Tomra has sorting devices available in the market that use NIR and visible spectroscopy, electromagnetic sensors and x-ray transmission[28]. They have used sensor fusion of

(24)

Figure 9. Results on hyperspectral imaging in the 400-1000nm region for contaminant detection[27].

color camera and electro magnetic sensor in their TITECH finder and TITECH com- bisense products. TITECH autosort is fusing NIR and VIS sensors. They also have product that uses either X-ray fluorescence or X-ray transmission combined with electro magnetic sensor. Figure 10 presents X-ray sorting. X-ray source below the conveyor belt creates X-ray radiation. When penetrating material and the radiation is attuanted X-ray camera measures with two independent sensor lines measures the response. This data is classified and the material is identified. They use pressurized air to sort classified rejects out of the material flow.

Figure 10.The working principle of Tomra’s XRT sorting with X-ray source and camera.

(25)

6 MACHINE VISION FOR PARTICLE SIZE DETEC- TION

In this chapter a machine vision system for particle size analysis is presented. Necessary steps for the system is presented. Methods that are used in the experiments are introduced.

Information is provided also about different segmentation methods.

6.1 Intraclass correlation

Intraclass correlation(ICC) is used to analyse the performance of the feature in this thesis.

In this particular case intraclass correlation is used for this purpose. ICC is a measure of reliability. In can be used to estimate the reliability of a of a feature. The intraclass correlation coefficient quantifies the propensity for observations on units within the same cluster to be more homogeneous than those in different clusters[29]. It describes how strongly units in the same group resemble each other. In this thesis ICC is utilized to analyse the correlation between particle size and segmentation result.

ICC is regarded in the framework of ANOVA (analysis of variance). The estimators used in this thesis are defined as [30]:

X_ij =µ+r_j+w_ij

X_ij is the ith observation of the jth group. µis the unobserved overall mean and r_j is random effect shared by all groupsw_ij is a residual normally distributed noise term. The variance ofr_j is denotedσ²_r and the variance ofw_ij is denotedσ_w² [31]. The ICCρis then defined as:

ρ= σ²_α σ_α² +σ²

6.2 Cumulative distribution function

Cumulative distribution function(CDF) is commonly used to model the distribution of particle sizes in RDF. CDF ofF(x) =P(X > x)describes the probability that random variableX takes has with given distributionF_X(x)to be less or equal tox. For example d₉₀value from the distribution means the such particle size that 90 percent of the particles

(26)

is smaller than d. Obtainingd_x value from the distribution is represented in Figure 11.

The properties of the CDF is that it is non-decreasing and continuous. In addition the distributionF(x)has property of

x→−∞lim F(x) = 0, lim

x→+∞F(x) = 1

That means thatd₁₀₀ value would be plus infinity. Function with these properties can be defined as a cumulative distribution. However, when distributions formed from empirical data Empirical cumulative distribution function (ECDF) are used. ECDF is not continuously as CDF. It is a step function with a saltus of 1/n and wheren is the amount of samples in the data and the d₁₀₀ value in cumulative histogram is corresponding to the largest particle in the measured data. Saltus means a jump in a discrete function.

0 10 20 30 40 50 60 70 80 90

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

d 90

particle size

cumulative percentage

Figure 11.Cumulative distribution andd₉₀value

6.3 Overview of the particle size detection system

To get the particle size detection system working it needs first to be trained. The training process from image acquisition to particle size distribution(PSD) has several steps:

1. Acquire images of RDF with known PSD 2. Preprocess the images

3. Segment the image with respect to the particles

4. Form cumulative distribution of segmented areas for each image 5. Least squares fit of the segmented areas to the ground truth distribution

(27)

In the first step acquired images needs to be labeled to their ground truth PSD. In the second step the images are preprocessed to reduce computing time in segmentation step. Also filter textures of the particles to improve segmentation. After that comes the segmentation of the particles and forming a cumulative distribution from the pixel areas of segmented regions. At this point the distributions created from each image. These distributions of segmented areas are the features used for and they are called in this thesis feature space.

The feature space is not the actual distribution of the particles but is the distribution that is computed from in the image. The final part of the process is to find a mapping from feature space to ground truth space. The segmentation and the fitting are the phases that needs the most research and are the most important for effective characterization of the particle size.

After the training image with unknown PSD can be used the with the same processing steps from 1 to 4. In the phase step 5 the fitted function can be used to map the cumulative distribution of segmented images to real sized particle distribution.

6.4 Image acquisition

For the experiments RDF sample has been obtained from the manufacturer. Manufacturer has provided such quality standards that particle size is smaller that 8x8x8 cm and one dimension is not bigger than 80 cm. The sample is divided manually to 3 different classes has that apparent difference in particle size. These samples are presented to camera shown in 12. Camera used in the experiments was a Canon EOS 450D. Halogen lamp was used to illuminate the RDF. Before acquisition of a picture the pile of RDF is shuffled to represent new image data from the same material. The resolution of the acquired image was 2736x3648 pixels. The number of pictures is not the defining factor of statistical significance but the amount of particles in the image and it is what we are interested in.

6.5 Determining ground truth for PSD

Determining the ground truth for each RDF class is needed for training and testing the machine vision system. Sieve analysis, which is the most commonly used method for particle size analysis, is not suitable for detailed analysis of RDF particles [25]. The heterogeneous nature of RDF makes the exact particle size analysis of the particle size a difficult task. Most of the provided particles of sample RDF are 2 dimensional particles

(28)

RDF

camera halogen

light

Figure 12.The setup for image acquisition of RDF samples.

such as paper and light plastics. Small amount of RDF is extracted from the samples.

The size distribution is determined by approximating the area of each particle from the extracted RDF. The average length and height of the particle is approximated and with meter manually. If the particle is a 3D object, the maximum projected area is considered.

6.6 Preprocessing

After acquiring the images is to be preprocessed. First the image is cropped to area where the waste fuel lies. After cropping the image is rescaled with ratio of 0.6 to reduce the computing time with out losing relevant information. To get rid of the most detailed texture Gaussian filtering is applied with a window size of 7x7 and with sigma value of 6.

After preprocessing the image resolution is reduced to 1249x2056.

6.7 Feature extraction

Getting the cumulative distribution from the image data seems to be difficult task. Fast segmentation of particles will oversegment while good segmentation of particles takes too much computing time to be applicable for online detection. Particles are overlapping and

(29)

there are textured, multicoloured, arbitrary shaped and even partially transparent particles.

This makes good segmentation of the particles impossible. However perfect segmentation is not needed for getting a nice estimate of the PSD. Instead of good segmentation, we should learn what a specific PSD segmentation looks like. In this thesis we have experimented segmentation with watershed transform, color clustering and region merging.

Also features extracted with Fourier transform are experimented.

6.7.1 Color clustering

K-means is a simple algorithm for cluster analysis. K-means takes an input vector of data points~xand the number of clustersK[32]. It returns an output vector~ywhich represents each data points assigned to one cluster. K-means algorithm goes as follows:

1. Make initial guesses forK amount cluster centroids 2. Attribute data points to the closest centroid

3. Move the cluster centroids to the mean of all data points belonging to that cluster 4. Repeat steps 2 and 3 until centroids no longer move

K-means clustering can be used for segmentation of the image by its color. Every pixel in the image is considered as a data point by their color. The location of the pixel is irrelevant. These data points are clustered with K-means clustering.

6.7.2 Statistical region merging

Statistical region merging(SRM) is a segmentation method that uses statistical basis for merging regions. It is used for remote sensing and medical image applications[33] [34].

SRM is based on the reconstruction of regions on the observed image based on unknown theoretical true image. I is an observation of a perfect sceneI∗that we do not know of.

The pixels in I∗are perfectly represented by a set of distributions, from which each of the observed color channel is sampled. Each color channel is replaced by a set of exactly Qindependent random variables. Controlling the scale of SRM segmentation is done by tuningQvariable. In 4-connective, there areN < 2|I|couples f adjacent pixels. LetS_I be a real valued function, withpandp⁰ pixels ofI. SRM first sorts the couples ofS_I in

(30)

Figure 13.Segmentation result by color clustering.

increasing order off(., .), and traverse this order only once. The for any current couple of pixels(p, p⁰)∈S_Ifor whichR(p)6=R(p⁰)the testP(R(p), R(p⁰)).R(p)stands for the current region R to which p belongs. MergeR(p)and R(p⁰) if the testP(R(p), R(p⁰)) returnstrue[35]. Figure 14 shows the segmentation result with SRM.

Figure 14.The segmentation result of SRM

(31)

6.7.3 Fourier transform

Fourier transform is used to change on image from the spatial domain to the frequency domain. The image is presented as a set of sinusoid functions. Fourier transform of a signalg(x, y)is defined as

F(g(x, y))(u, v) =

∞

Z Z

−∞

g(x, y) exp−i2π(ux+vy)dxdy

[36]. The process takes complex valued imagesg(x, y)with zero imaginary component.

Fourier transform output consist of phase and magnitude. Figure 15 shows phase and logarithmic magnitude of fuorier transrom. The real value is the magnitude of sinusoid and the complex part is represents the phase. It is assumed that the images containing smaller particle size RDF has higher magnitude high frequencies in the fourier image.

associate with faster changes in the image. Fourier transform is applied to preprocessed images to estimate mean particle size of the image. In this experiment the ratio of low and high amplitudes in the Fourier image is used as a feature for mean particle size. The lowest frequencies are the pixels in the Fourier image that are in the center inside a circle of 6 pixel radius. Figure 16 shows an example image that is decomposed into low and high frequencies accordingly.

(a) (b) (c)

Figure 15. Fourier transform split to low and high frequencies (a) Fourier transform logarithmic magnitude (a) Fourier transform logarithmic magnitude (b) Fourier transform phase.

6.7.4 Watershed transform

Watershed transform is one method performing image segmentation originated from [37].

It is used for defining particle size distribution in river beds [38] [39]. The watershed transformation considers the image gradient magnitude as topographic surface where gray level of a pixels is interpreted as its altitude. Pixels having the highest gradient magni-

(32)

(a) (b)

Figure 16.Image split to low and high frequencies (a) Low frequency image. (b) High frequency image.

tude intensities are considered as watershed lines, which represents region boundaries. A drop of water falling on topographic relief flows along a path to local minimum. Pixels draining water to a common minimum are representing a segment. Figure 18 illustrated how watershed transform segments are formed. From each dotted line the water will be draining to a local minimum. Figure 17b illustrated how to avoid oversegmentation by suppressing insignificant minimums. The segmentation is demonstrated in figure 18.

watershed regions

(a)

M1 M2 M1

Watershed regions

(b)

Figure 17. Watershed transform illustrated in one-dimensional case (a) Watershed segmentation;

(b) Watershed segmentation with minimum suppress.

6.7.5 Superpixels

Superpixels algorithms groups pixels in perceptually meaningful pixel areas rather than rigid grid of pixels [40]. The simple linear iterative clustering(SLIC) algorithms is one of the state-of-the-art method which adapts k-means clustering to generate superpixels in

(33)

(a) (b)

Figure 18.Watershed transform result (a) Original image; (b) Watershed transform result.

similar way as [41]. Figure 19 illustrates the how image is segmented with SLIC superpixels. Superpixels has a parameter predefined parameter which approximately defines the amount of segments in the image. There for it is not by it self application for defining particle size but may be useful as pre-processing.

(a) (b)

Figure 19.Superpixels segmentation with superpixels. (a) Original image; (b) SLIC superpixels

(34)

7 EXPERIMENTS AND RESULTS

This chapter describes the datasets and the experiments done in the thesis. Firstly The datasets are presented. Secondly tests and their results for choosing the most suitable tools for particle size analysis. Finally the results on particle size analysis are presented.

7.1 Datasets

The data for the first experiment consist of total of 48 images acquired from refuse derived fuel sample. We have acquired 16 images 3 different classes that have different particle size. Figure 20 illustrates the size difference of the particles showing one example from each class.

Figure 21 shows the cumulative distribution of each separated class of particle size. Red line is the empirical cumulative distribution(ECD) of the most bulkiest RDF in Figure 20a, green is the ECD of medium size RDF in figure20b and blue is ECD of the small size RDF in Figure 20c. ECD of each is defined by manually measuring the particles and forming a cumulative distribution of the results. These classes haved₅₀value of 50 cm of the bulkiest RDF, 17 cm of middle size RDF and 7 cm of the small size RDF.

For testing purposes we produced also a dataset with different illumination using the camera flash of Canon EOS 450D. There are 8 pictures per class with the same ground truth as halogen light. For the mixed dataset, we mixed the previously known PSD to create new ones. Firstly, small and medium class has been mixed with equal volume from both classes. Similarly we mixed small and bulk class and also all classes in together.

The resulting PSD would be therefore a combination of the mixed PSD and the average particle size would be in between the classes where the mixture is originated from. Also 8 pictures of each mixed class has been taken using the halogen light. All the datasets are displayed in Table 2.

7.2 Initial test

The experiments were done in 2 phases. In the first phase the objective was to be able to separate each class from the other. In the second phase the cumulative size distribution of the RDF is approximated from the image data. Figure 22 shows features which

(35)

(a) (b) (c)

Figure 20. Example images of (a) bulk, (b) medium and (c) small particle size of RDF with halogen illumination.

Table 2.Datasets

Dataset Particle size distribution

Halogen light small medium bulk

N 16 16 16

Camera flash small medium bulk

N 8 8 8

Mixed dataset small and small and small, medium With halogen light medium bulk and bulk

N 8 8 8

were extracted with Fourier and the watershed transform. Each class corresponding to one particle size is distinct in the feature space. It means that different particle size is rec- ognizable and the used features has at least some correlation to particle size of the RDF.

The next step is to define the ground truth PSD.

To find out a good feature for PSD estimation some segmentation and feature extraction methods are considered. To estimate the performance of the feature extracting methods ICC is employed. Table 3 shows the ICC of color clustering, region merging, Fourier transform and watershed transform features. The feature for each at this point is the number of segmented areas from each image. The methods that have the highest ICC value are studied further and employed for defining the particle size distribution.

The ICC value roughly represents the probability of the sample to be classified to the

(36)

0 50 100 150 200 250 300 0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Particle size (cm²)

Cumulativepercent(%)

bulky medium small

Figure 21.Ground truth of the three different distributions.

correct class. Every method uses preprocessed images. Color clustering clusters pixel values to 8 different class with K-means. Then it calculates the amount of 8 connected regions in the picture. Region merging is done with Q = 512. Watershed transform method has the following pre steps:

1. Transform the image to grayscale image 2. Median filter the grayscale image

3. Find the image gradient with sobel operator 4. Perform minimum suppression

First after preprocessing the color image is changed to grayscale image. Median filter size of 5x5 is used to remove color and texture variation inside the particles which results in filtered grayscale image. Then image gradient is magnitude is calculated with sobel filter which turns the grayscale image to edge image. To reduce over segmentation of the watershed transform minimum filter is used to suppress insignificant minimums. Finally from the minimum suppressed edge image the watershed transform is done to segment the image.

In Figure 22 two best features are displayed the information based on the two best features , Fourier transform and watershed transform in two dimensional plot. The three different classes, bulk medium and small forms an own cluster. That means that it is possible

(37)

Table 3.ICC of different feature extraction methods Feature extraction method ICC

Color clustering 0.99393

Region merging 0.99425

Fourier transform 0.99833 Watershed transform 0.99907

to distinguish different classes from each other by machine vision. The axes shows the amount of segmented particles in the image. There is no information about PSD.

300 350 400 450 500 550 600 650 700

0.945 0.95 0.955 0.96 0.965 0.97 0.975

Watershed feature

Fourier feature

bulky medium small

Figure 22.Watershed and Fourier features of different particle size RDF

7.3 Particle size distribution

Watershed transform is utilized to define the particle size distribution with the same steps as in initial test.

(38)

Figure 23 shows the distributions of watershed areas. There is 8 image for each class from the dataset with halogen light. Each line represents distribution of segmented areas of one image. Classes are separated in the distribution extracted from the segmentation but the shape of the curve is not similar to the ground truth in figure 21. The problem is to find a suitable mapping from the feature space to the ground truth space. It this study linear first degree polynomial is used. At this phase it is trained what segmentation of a specific PSD looks like. Let be a vector defineX~ the cumulative distribution in feature space, and let Y~ be a same size vector in ground truth space, is also the desired output.

The goal is to find a model that fitsf(X) =~ Y~. Now the fitting is done as a linear model:

f_i(X_i) =a_iX_i+b_i. Figure 24 shows how the the linear model maps the PSD from pixel real PSD. On the left side PSD is obtained from the segments of image segmentation and on the right side there is a PSD that the function estimates as its ground truth PSD.

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x 10⁴ 0

10 20 30 40 50 60 70 80 90 100

Pixel Area

Cumulative Area %

Figure 23.Distribution from watershed segmentation

7.4 The Results from the mapping

For eachd_i value from the feature space was fitted linearly in least square sense to thed_i value of the ground truth space. The result of the fitting is shown in figure 25 and more closely on small and medium classes in 26. Colors represent each class of different PSD.

Each dotted line equals to one distribution acquired from the image meaning the color distribution observed by machine vision. Thick lines represents the ground truth of their color.

(39)

Particle size in pixels

f₁₀₀ x₁₀₀

f₁ x₁

x_i f_i

x_i y_i

Particle size in cm²

Cumulaive percentage

%

Cumulaive percentage

%

Figure 24.Illustration of the mapping of particles from features space to ground truth space.

In figure 27 different PSD of 3 mixed classes measured by machine vision are shown with using different illumination than in the training. In this test we are using camera flash instead of halogen light in the training. Inequality in the illumination has dislocated slightly the measured PSD.

In figure 28 different PSD of 3 mixed classes measured by machine vision are shown.

There is 8 PSD calculated from 8 image from each class. The PSD of mixture of small and medium particles shows very reasonable result. The resulting PSD is between the Ground truths. The results of mixture of small medium and bulk and small bulk are also in reasonable results. However the proposed machine vision system cannot distinct these classes from each other. The method was implemented in MATLAB, using a PC with a 3.20 GHz CPU and 8 GB of RAM. The processing of defining the PSD took less than half second per acquired image.

(40)

0 50 100 150 200 250 300 350 400 0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

small GT med GT bulk GT small measurement med measurement big measurement

Figure 25.Cumulative PSD obtained with machine vision.

0 10 20 30 40 50 60 70 80 90 100

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 26.Cumulative PSD obtained with machine vision small and medium classes.

(41)

0 50 100 150 200 250 300 350 400 0

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 27.Cumulative PSD with different illumination than used in training.

0 50 100 150 200 250 300

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

small GT med GT bulk GT small med small med bulk small bulk

Figure 28. PSD of mixed classes obtained with machine vision. Legend shows the what classes are mixed with equal volumes.

(42)

8 DISCUSSION

8.1 Machine vision in RDF manufacturing

The purpose of this thesis was to investigate possibilities of machine vision system in the quality assurance of RDF. Utilization of machine vision in manufacturing process or in quality assurance of RDF process is quite new in the field. There were only few scientific papers on imaging techniques in RDF process and only few products available in the market.

8.2 Getting the particle size distribution

Defining the particle size with a machine vision turned out to be a difficult task because of the challenges of segmenting waste fuel particles. The shape of the distributions of the material and segmented particles were not similar. One subtask was to find the most suitable mapping from the segmented distribution to the ground truth distribution. The most innovative part of the particle detection is the fitting from the feature space to the ground truth space. The fitting equals to the training what the segmentation of specific PSD looks like. The correlation between the segmentation and the PSD is described in the parameters of the vector of linear functions. However, the real correlation between real PSD and the image segmentation is far more complex. It is originated from the overlapping of objects, oversegmentation, attributes of the material shape and texture and numerous other nature of a pile of RDF.

8.3 The results

In this work PSD of RDF material is defined with a machine vision system. The results with the same conditions seems to be quite accurate. However, the use of the same PSD in the training and testing suggest better results than the system is capable of. Using slightly different illumination does not seem to have a huge effect on the results. The result of mixed classes are not as accurate because such classes are not introduced in the teaching data and the variance inside the class is bigger and therefore more particles and images needs to be analysed so that the PSD can be defined accurately. These mixed classed can not be distinguished based on one picture only. However the results of mixed class RDF

(43)

makes sense and the PSD makes sense since the obtained PSD is between the classes where the RDF sample is originated from. The ground truth of these samples were not acquired. It is easy to acquire more data than only one image in a RDF manufacturing plant of RDF material. The more data and images are taken from the data the more accurately the material can be characterized.

8.4 Future Work

The best results are received using watershed transform which uses gray scale images which means that color information has not been used to improve performance. The limitations of the PSD should be tested with more particle size distribution classes with different variances inside the size distributions. Different kind of segmentation supposed to be tested for example using clustering of superpixels. The mapping of segmented areas to the ground truth of the PSD needs more research and more data. More labeled classes images with ground truth needs to be done to achieve better performance of the system.

By this method is should be tested how accurately it is possible to describe the material using tens or hundreds of images.

(44)

9 CONCLUSIONS

Machine vision has become useful tool for RDF manufacturing process. Hyperpectral, NIR and X-ray cameras are already used for recycling and increasing the biomass fraction of RDF end product. NIR and X-ray fluorescence spectrometry is used to determine the composition RDF. Grayscale and RGB cameras have been used for determining the particle size of RDF. There are also machine solutions to characterize RDF or SRF in lab- oratory conditions by sampling. However, in this thesis on-line machine vision systems are study to characterize waste fuels. Characterize in realtime in the middle of the process makes the task more challenging and possibility to adjust automation process and to detect malfunctions. A good particle size analysis in this work has not come from a good segmentation of particles that would extract every particle in the image correctly. It has come from by learning the segmentation results for images with different PSD. Distinc- tiveness is measured with ICC. The meaning segmentation result is learned to represent a specific PSD. The proposed system for particle size distribution analysis shows promising results and the implementation of such machine vision system is feasible.

(45)

REFERENCES

[1] Cen tc 343 - "solid recovered fuel". http://erfo.info/cen-tc-343.26.

0.html. Accessed: October 2014.

[2] R Sarc and Karl E Lorber. Production, quality and quality assurance of refuse derived fuels (rdfs). Waste management, 33(9):1825–1834, 2013.

[3] Antonio Gallardo, Mar Carlos, MD Bovea, Francisco J Colomer, and Fernando Al- barrán. Analysis of refuse-derived fuel from the municipal solid waste reject fraction and its compliance with quality standards. Journal of Cleaner Production, 83:118–

125, 2014.

[4] Emmanuel Kakaras, Panagiotis Grammelis, Michails Agraniotis, Willy Derichs, Hans-Peter Schiffer, Jörg Maier, Thomas Hilber, Thomas Glorius, and Uwe Becker.

Solid recovered fuel as coal substitute in the electricity generation sector. Thermal Science, 9(2):17–30, 2005.

[5] Petr Stehlík. Contribution to advances in waste-to-energy technologies. Journal of Cleaner Production, 17(10):919–931, 2009.

[6] Masaru Tanaka. Waste management for a sustainable society. Journal of Material Cycles and Waste Management, 9(1):2–6, 2007.

[7] European Parliament and Council of the European Union. Directive 2000/76/EU, 2000.

[8] European Parliament and Council of the European Union. Directive 2010/75/EU, 2010.

[9] Wenchao Ma and Susanne Rotter. Overview on the chlorine origin of msw and cl- originated corrosion during msw & rdf combustion process. InBioinformatics and Biomedical Engineering, 2008. ICBBE 2008. The 2nd International Conference on, pages 4255–4258. IEEE, 2008.

[10] Sabine Flamme and Julia Geiping. Quality standards and requirements for solid recovered fuels: a review. Waste Management & Research, 30(4):335–353, 2012.

[11] Karl E Lorber, Renato Sarc, and Alexia Aldrian. Design and quality assurance for solid recovered fuel. Waste Management & Research, 30(4):370–380, 2012.

[12] CA Velis, Philip J Longhurst, Gillian H Drew, Richard Smith, and Simon JT Pol- lard. Production and quality assurance of solid recovered fuels using mechanical

(46)

biological treatment (mbt) of waste: a comprehensive assessment. Critical Reviews in Environmental Science and Technology, 40(12):979–1105, 2010.

[13] C Eswaraiah, T Kavitha, S Vidyasagar, and SS Narayanan. Classification of metals and plastics from printed circuit boards (pcb) using air classifier. Chemical Engi- neering and Processing: Process Intensification, 47(4):565–576, 2008.

[14] M Shapiro and V Galperin. Air classification of solid particles: a review. Chemical Engineering and Processing: Process Intensification, 44(2):279–285, 2005.

[15] Md Abdur Rahman and MCM Bakker. Sensor-based control in eddy current separation of incinerator bottom ash. Waste management, 33(6):1418–1424, 2013.

[16] Ruan Jujun, Qian Yiming, and Xu Zhenming. Environment-friendly technology for recovering nonferrous metals from e-waste: Eddy current separation. Resources, Conservation and Recycling, 87:109–116, 2014.

[17] Ren C Luo, Chih-Chen Yih, and Kuo Lan Su. Multisensor fusion and integration:

approaches, applications, and future research directions. Sensors Journal, IEEE, 2(2):107–119, 2002.

[18] Liu Bin and Peng Jiaxiong. Image fusion method based on nonseparable wavelets.

Machine Vision and Applications, 16(3):189–196, 2005.

[19] Tako PR DE JONG, JA van Houwelingen, and W Kuilman. Automatic sorting and control in solid fuel processing: opportunities in european perspective. Geologica belgica, 7(3-4):325–333, 2004.

[20] Tuomas J Lukka, Timo Tossavainen, Janne V Kujala, and Tapani Raiko. Zenrobotics recycler–robotic sorting using machine learning.

[21] Ian T Young, Jan J Gerbrands, and Lucas J Van Vliet. Fundamentals of image processing. Delft University of Technology Delft, The Netherlands, 1998.

[22] Norbert Huber, Simon Eschlböck-Fuchs, H Scherndl, A Freimund, J Heitz, and JD Pedarnig. In-line measurements of chlorine containing polymers in an industrial waste sorting plant by laser-induced breakdown spectroscopy. Applied Surface Science, 302:280–285, 2014.

[23] Petra Tatzer, Markus Wolf, and Thomas Panner. Industrial application for inline material sorting using hyperspectral imaging in the nir range. Real-Time Imaging, 11(2):99–107, 2005.

(47)

[24] W KUILMAN, JA van HOUWELINGEN, et al. Automatic sorting and control in solid fuel processing: Opportunities in european perspective. Geologica Belgica, 2006.

[25] Gregory Dunnu, Thomas Hilber, and Uwe Schnell. Advanced size measurements and aerodynamic classification of solid recovered fuel particles. Energy & fuels, 20(4):1685–1690, 2006.

[26] Artzai Picon, Ovidiu Ghita, Pedro M Iriondo, Aranzazu Bereciartua, and Paul F Whelan. Automation of waste recycling using hyperspectral image analysis. In Emerging Technologies and Factory Automation (ETFA), 2010 IEEE Conference on, pages 1–4. IEEE, 2010.

[27] Silvia Serranti, Aldo Gargiulo, and Giuseppe Bonifazi. Characterization of post- consumer polyolefin wastes by hyperspectral imaging for quality control in recycling processes. Waste Management, 31(11):2217–2227, 2011.

[28] Tomra sorting devices. http://tomra.com/en/. Accessed: October 2014.

[29] Gary G Koch. Intraclass correlation coefficient.Encyclopedia of statistical sciences, 1982.

[30] Kenneth O McGraw and Seok P Wong. Forming inferences about some intraclass correlation coefficients. Psychological methods, 1(1):30, 1996.

[31] William M Stanish and Noel Taylor. Estimation of the intraclass correlation coefficient for the analysis of covariance model. The American Statistician, 37(3):221–

224, 1983.

[32] James MacQueen et al. Some methods for classification and analysis of multivari- ate observations. InProceedings of the fifth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281–297. Oakland, CA, USA., 1967.

[33] M Emre Celebi, Hassan A Kingravi, Hitoshi Iyatomi, Y Alp Aslandogan, William V Stoecker, Randy H Moss, Joseph M Malters, James M Grichnik, Ashfaq A Marghoob, Harold S Rabinovitz, et al. Border detection in dermoscopy images using statistical region merging. Skin Research and Technology, 14(3):347–353, 2008.

[34] Zhijian Huang, Jinfang Zhang, Xiang Li, and Hui Zhang. Remote sensing image segmentation based on dynamic statistical region merging. Optik-International Journal for Light and Electron Optics, 125(2):870–875, 2014.

[35] Richard Nock and Frank Nielsen. Statistical region merging. Pattern Analysis and Machine Intelligence, IEEE Transactions on, 26(11):1452–1458, 2004.

(48)

[36] David A. Forsyth and Jean Ponce. Computer Vision: A Modern Approach. Prentice Hall Professional Technical Reference, 2002.

[37] Fernand Meyer and Serge Beucher. Morphological segmentation. Journal of visual communication and image representation, 1(1):21–46, 1990.

[38] MS Sulaiman, SK Sinnakaudan, SF Ng, and K Strom. Application of automated grain sizing technique (ags) for bed load samples at rasil river: A case study for supply limited channel. Catena, 121:330–343, 2014.

[39] Fi-John Chang and Chang-Han Chung. Estimation of riverbed grain-size distribution using image-processing techniques. Journal of Hydrology, 440:102–112, 2012.

[40] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk. Slic superpixels compared to state-of-the-art superpixel methods.

Pattern Analysis and Machine Intelligence, IEEE Transactions on, 34(11):2274–

2282, 2012.

[41] C Lawrence Zitnick and Sing Bing Kang. Stereo for image-based rendering using image over-segmentation. International Journal of Computer Vision, 75(1):49–65, 2007.

Quality assurance of refuse derived fuel with machine vision

QUALITY ASSURANCE OF REFUSE DERIVED FUEL WITH MACHINE VISION

CONTENTS

ABBREVIATIONS AND SYMBOLS

1 INTRODUCTION

1.1 Background

1.2 Objectives and Restrictions

1.3 Structure of the Thesis

2 QUALITY ASSURANCE OF REFUSE DERIVED FUEL

2.1 RDF as energy

2.2 Quality criteria of waste fuels

2.3 Sampling

2.4 Characterization of waste fuels

3 RDF PRODUCTION

3.1 Typical steps of RDF production

3.2 Size reduction

3.3 Ballistic separation

3.4 Mechanical screening

3.5 Air classification

3.6 Magnetic separation

3.7 Eddy current separation

4 SENSOR FUSION

4.1 Low-level fusion

4.2 Medium-level fusion

4.3 High-level fusion

4.4 Sensories used in RDF process

5 MACHINE VISION AND RDF PROCESS

5.1 Introduction to machine vision

5.2 Benefits of machine vision in RDF production

5.3 Existing machine vision solutions

6 MACHINE VISION FOR PARTICLE SIZE DETEC- TION

6.1 Intraclass correlation

6.2 Cumulative distribution function

6.3 Overview of the particle size detection system

6.4 Image acquisition

6.5 Determining ground truth for PSD

RDF

6.6 Preprocessing

6.7 Feature extraction

7 EXPERIMENTS AND RESULTS

7.1 Datasets

7.2 Initial test

7.3 Particle size distribution

7.4 The Results from the mapping

8 DISCUSSION

8.1 Machine vision in RDF manufacturing

8.2 Getting the particle size distribution

8.3 The results

8.4 Future Work

9 CONCLUSIONS

REFERENCES